Scoped memory for coding agents. A router in front of two stores.
question ──▶ router ──▶ ┌── structure store what calls what
(no LLM) └── episode store what happened and why
└─▶ merged, cited answer │
or ABSTAIN ────────────┘
Knowledge graphs hold structure but no prose, so they cannot answer . Verbatim-recall systems hold prose but no call graphs, so
they cannot answer . Both make you name the
namespace when you .
Scoped memory for coding agents: picks the right project, answers in scope, or abstains.
2026-11-08
Sao
—
Công cụ
—
Thời gian hoạt động
—
why did the
cookie get dropped
what calls run_agent_turn
write
loci decides the scope when you read. One question path serves "how does
auth work here" and "have I solved this in any project" — only the size of
the scope set changes.
The shape, in one picture
A street of independent shops. Each shop is one project — its own stock, its
own books, nothing pooled in a shared warehouse. Each keeps two ledgers:
what calls what, and what happened and why. They are read together or not at
all.
Shops wear overlapping badges — yours, this client's, that vendor's, this
monorepo's. A badge can cover shops at opposite ends of the street, which is why
they are labels rather than a hierarchy.
You never walk the street yourself. You ask the dispatcher, who knows only
which uncommon words are stocked where. They send you into at most three
shops, or say nobody here carries that and name who might.
Each shop also hangs a signboard naming what it is — repository, package,
command — and keeps a list of what it reaches for. Reading one against the
other is how the street answers which of my shops buys from another, which no
amount of asking the dispatcher ever could: that question is about the road
between two shops, and the dispatcher only knows what is stocked inside them.
One thing the street still lacks: aisles inside a shop, so a word found in a
test fixture does not count like the same word in shipped code.
The problem it solves
Put every project in one index and the largest one wins regardless of the
question. Measured against a merged graph of ten real repositories:
question
on-topic nodes returned
"why was the admin session cookie dropped on localhost?"
18% — 19 of 31 came from the biggest project
"what happens when a user clicks Save as ZIM?"
2% — 61 of 62 came from the biggest project
"how does the reserved landing step work?"
98% — and the biggest project was the answer
That third row is the trap. Merged retrieval looks excellent whenever the answer
happens to live in the largest corpus and collapses when it does not. It is the
same failure in both directions, visible in only one.
Install
The distribution is loci-mem; the command, the import and the project are
all loci. PyPI's loci is an unrelated outlier-detection package abandoned in
2018 — the same split as python-dateutil installing as dateutil.
pipx install loci-mem # routing + lexical search, one dependency
pipx install 'loci-mem[all]' # + graphify, local embeddings, MCP server
The base install pulls numpy, and nothing else. Routing, BM25 and the
char-gram matrix are a compiled Rust extension as of 0.6.0, so scikit-learn
(and scipy behind it), rank-bm25 and joblib are gone. No torch, no vector
database, no model download unless you ask for one.
Wheels are published for macOS, Linux and Windows; installing from the sdist
builds from source and needs a Rust toolchain (1.75+).
extra
adds
graphify
code-symbol extraction across ~18 languages
embeddings
local bge-small semantic ranking
rerank
cross-encoder reranking (opt-in per query)
mcp
the MCP server
Quickstart
pipx install 'loci-mem[all]'
loci setup # scan, graph, index, embed, calibrate -- one pass
loci ask "why was the session cookie dropped on localhost?"
setup asks only what it cannot decide for you: which directories hold your
projects, whether to register the repositories it found that are not yours, and
whether to spend a one-time model download on semantic search. It ends by
running doctor, so whatever it could not cover is the last thing you read
rather than something you discover from a bad answer a week later.
Every prompt takes its default when stdin is not a terminal, so it is safe to
run unattended in a container or under an agent. -y does the same from a
terminal, and --no-graphs / --no-embed / --no-calibrate decide individual
steps up front.
The same thing by hand. The order is a dependency chain, not a preference:
graphs are what the index is built from, the index writes the chunks embed
encodes, and calibrate fits its semantic floor from those vectors.
loci scan ~/code # register every git repo it finds, one scope each
loci graphs # optional: add code symbols (free, no model calls)
loci index # build the routing index + episode store
loci embed # optional: local vectors for semantic recall
loci calibrate # optional: fit routing thresholds to your corpus
loci doctor # what is missing, and the command that fixes it
loci ask "which projects use wrangler and D1?"
loci eval # measure routing accuracy on YOUR corpus
Each step writes what the next one reads, which is what makes the order fixed
rather than preferred:
flowchart TD
scan["loci scan ~/code<br/>writes ~/.loci/scopes.json"]
graphs["loci graphs<br/>writes graphify-out/graph.json, one per scope"]
index["loci index<br/>writes ~/.loci/episodes.json, scope_index.json, rankers/"]
embed["loci embed<br/>writes ~/.loci/embeddings.npz"]
calibrate["loci calibrate<br/>writes ~/.loci/calibration.json"]
doctor["loci doctor<br/>writes nothing at all"]
scan -->|"there is nothing to graph or index until a scope exists"| graphs
graphs -->|"the index is built FROM graph.json; index first<br/>and every project routes on prose alone"| index
index -->|"embed encodes episode chunks, and they do not<br/>exist until index has written the store"| embed
embed -->|"calibrate fits a per-scope semantic floor from<br/>those vectors; with none it keeps a default"| calibrate
calibrate -->|"and then reports whatever is still missing"| doctor
Running it out of order does not fail loudly. It produces an install that works
and quietly retrieves worse, which is the failure doctor exists to name.
scan registers one scope per git repository. It also reads who owns each
repository out of git, prints who owns what, and asks before registering the
ones that are not yours; that prompt takes its default like every other, which is to
register everything.
A monorepo can become one scope per package instead. --split on loci scan or
loci setup takes anything carrying package.json, pyproject.toml,
Cargo.toml or go.mod one level down; a repo-local .loci.json names the rest
and is honoured with or without the flag. --split is off by default, and
the reason is measured rather than cautious: a new scope's aliases include its
bare directory name, and an alias outranks your working directory (6.0 to 4.0).
On the development corpus, splitting a repository holding glasses/ sent eight
hand-written questions about a different project to Delroy/glasses — seven of
them had routed correctly before the split, and six of the eight reverted when
the alias boost was zeroed. A Jekyll _site/ build directory became a scope of
its own on the same run. A .loci.json you wrote does not have that problem: you
named those sub-projects deliberately, so you can see what they are called.
ask uses your working directory by default, and should. Questions that name no
project — "how is this deployed?", "how do I run the tests?" — route
correctly 100% of the time with cwd. Without it they are unanswerable, and
loci abstains on all of them rather than guessing. cwd is not a tiebreaker
signal here, it is the primary one.
Concepts
Scope. One project, one namespace. A registered git repository, one package
of a monorepo (under --split, or named in .loci.json), or anything you add
explicitly. Scopes are never merged.
Group. An overlapping label on a scope: me and vendor:<org>, read from
git provenance; client:acme and anything else you assert by hand; and a
monorepo's own id, carried by every package inside it and by the monorepo
itself, so --group <monorepo> reaches the code no package claimed. A scope can
be in several, and the scope set stays flat: grouping never merges scopes or
nests one inside another.
What a group does to a question is its mode, and the mode answers two
different questions. Reached through your working directory, explicit does
nothing, soft (the default) multiplies every outside scope's evidence base by
0.5, and hard confines routing to members and abstains when the best answer is
outside. Named on the command line with --group X, all three modes confine
to X's members — the mode decides only what happens when the best answer is
outside them: hard abstains and says so, soft and explicit answer with the
best member anyway. Membership lives in the scope registry, mode in
groups.json, so a re-scan — which rewrites the registry wholesale — cannot
discard policy. Measured: hard
anchored on cwd fires on questions that name an outside project (12 of 12) and
not on questions carrying only its vocabulary (0 of 24), so in practice --group
and project names drive it rather than where you are standing.
Structure store. What calls what — symbols, files, references, traversals
with file:line citations. Supplied by graphify
through an adapter.
Episode store. What happened and why — README and docs, git commit bodies,
docstrings and comment blocks, and any notes you point it at. Stored verbatim,
chunked on heading boundaries, redacted before it is written.
Router. Decides which scopes a question belongs to, deterministically, with
no model call. Signals in order of weight: an explicit project name (6.0), the
working directory (4.0), vocabulary evidence (measured 0.1–1.5), recency (0.15).
Weight is not usefulness — cwd is the signal that carries most real questions,
because most real questions name no project at all.
Abstention. A first-class outcome. ABSTAIN means ask the user, not pick
the biggest. Both layers can refuse.
An abstention that only says no leaves the asking to you, so it carries the
shortlist it refused to choose between — each project, and the terms that put it
there:
loci ask "how is caching handled?"
ABSTAINED - not enough of the question exists in any project.
candidates: Delroy (handled, caching), odysseus (handled, caching), tensor-serve (caching)
re-run with --scope <name>, or from inside the project directory.
A scope earns its place by holding a term at most half the corpus holds.
Holding one is not enough: change, handled and work sit in nearly every
project, so matching them is a coincidence rather than a claim. Measured on the
development corpus, that is the difference between naming three projects and
listing all fourteen in score order, which is the registry rather than a
shortlist.
When nothing clears that bar there is no shortlist to print, and saying so is the
answer — the question's subject is not indexed anywhere, and choosing a scope
would only relocate the guess:
loci ask "xyzzy plugh frotz quuz"
ABSTAINED - not enough of the question exists in any project.
no project holds a distinctive term from this question -- `loci doctor` shows what is not indexed.
Architecture
The query path
The diagram at the top of this file is the elevator version. In full:
loci ask "why was the session cookie dropped on localhost?"
│
│ --scope NAME jumps straight past routing to the fan-out, and drops the
│ episode gate with it: that gate exists to stop an answer arriving from
│ the wrong scope, and you have just named the right one.
▼
confinement groups.py
│ reads the registry and groups.json. Never the index, never a model.
│ --group X ─▶ eligible = members of X, in ALL THREE modes
│ else cwd ─▶ the strictest group of the scope you are standing in
▼
route router.py
│ reads scope_index.json: token ─▶ {scope: node_df}. One dict lookup per
│ query token. Deterministic, sub-millisecond, no model call.
│
├──▶ ABSTAIN — deictic, no_evidence or out_of_group. Names the cause, lists
│ the scopes with a claim on the question and what each one
│ holds, names the flag that fixes it, and queries nothing.
│ An outcome, not an error.
▼
selected scopes, at most 3 one thread each; neither store is ever
│ queried across a scope boundary
├─── scope ─── scope ─── scope
│ │
│ │ expand the question against THIS scope's postings — a dict
│ │ lookup, so a token the scope does not have cannot be invented —
│ │ then append the tokens of its nearest embedded symbol labels
│ │
│ ├─ structure store graphify query --graph <this scope's graph>
│ │ what calls what, with file:line citations
│ │
│ └─ episode store BM25 + char 3–5 gram + embeddings, fused
│ gated: lexically grounded OR semantically
│ confident, so it can return nothing, and does
▼
merged, cited answer — one block per scope, and `no evidence in this scope`
wherever both of a scope's stores came back empty
Both stages refuse independently: the router can decline to pick a scope at
all, and a scope it did pick can still hand back nothing.
The routing decision
Everything above the fan-out is one function. It scores every scope in the
corpus, then asks three refusal questions in a fixed order. An alias or cwd
signal skips the last two, but not the first.
flowchart TD
Q["question, plus cwd and --group when given"]
CONF{"which group confines this question?"}
E1["eligible = members of X<br/>strict only when that group's mode is hard"]
E2["eligible = members<br/>strict"]
E3["demoted = every scope outside the group"]
E4["nothing confined, nothing demoted"]
BASE["evidence base, for EVERY scope in the corpus<br/>sum(scope-idf x prominence) / query tokens / size^0.15"]
PEN["demoted? base x 0.5<br/>here, before the boosts, so it cannot invert one"]
BOOST["+ ALIAS 6.0, the question names the project<br/>+ CWD 4.0, DEEPEST containing scope only<br/>+ RECENCY 0.15, a tiebreak and no more"]
TOPS["top = best ELIGIBLE scope<br/>top_all = best scope in the whole corpus"]
OOG{"strict, and top_all is outside the group,<br/>and top_all would itself have been routable?"}
A1(["ABSTAIN out_of_group<br/>the best answer is outside the group"])
FORCED{"forced?<br/>top carries an alias hit or the cwd signal"}
ENUM{"enumerative?<br/>which projects / what repos / any of my / where else / have I ever"}
DEIX{"deictic?<br/>this / these / it / its / here / the project / the app / ..."}
A2(["ABSTAIN deictic<br/>the question points at a subject it never names"])
EV{"enough evidence?<br/>ANY ONE of the three is enough"}
A3(["ABSTAIN no_evidence<br/>too little of the question exists in any project"])
SEL{"did the question ask for a SET?"}
SELECT["keep every eligible scope scoring at or above 0.85 x top,<br/>cap at 3, then force in every concentrated-token holder"]
SETSEL["keep every eligible scope clearing 0.8 x the floor ON ITS OWN,<br/>in rank order, cap at 8 -- no ratio cutoff at all"]
OUT(["selected scopes"])
Q --> CONF
CONF -->|"--group X, any mode"| E1
CONF -->|"cwd scope's strictest group is hard"| E2
CONF -->|"cwd scope's strictest group is soft"| E3
CONF -->|"explicit, or no group at all"| E4
E1 --> BASE
E2 --> BASE
E3 --> BASE
E4 --> BASE
BASE --> PEN
PEN --> BOOST
BOOST --> TOPS
TOPS --> OOG
OOG -->|yes| A1
OOG -->|no| FORCED
FORCED -->|yes| SEL
FORCED -->|no| ENUM
ENUM -->|"yes, and it is not picking ONE scope"| EV
ENUM -->|no| DEIX
DEIX -->|yes| A2
DEIX -->|no| EV
EV -->|"summed token evidence at or above the floor<br/>(7.6 shipped, refitted by loci calibrate)"| SEL
EV -->|"at least 4 matched tokens"| SEL
EV -->|"an eligible scope holds a CONCENTRATED token:<br/>held by at most 2 scopes, prominent inside them"| SEL
EV -->|"none of the three"| A3
SEL -->|no| SELECT
SEL -->|yes| SETSEL
SELECT --> OUT
SETSEL --> OUT
Four things in that shape are load-bearing and none of them are obvious.
forced is an escape, not a signal. An alias hit or a cwd hit contributes
nothing to either evidence count, so a scope winning purely on 6.0 or 4.0 reads
as zero evidence to the gates below. It skips both of them instead: deixis is
only a problem when nothing else has identified the subject, so a question that
says "it" and names a project is unaffected.
out_of_group is judged against the whole corpus, not the group. It fires
only when the corpus-wide winner is outside the group and would itself have
routed. Without that second test it swallowed the other two reasons entirely —
on a question matching no vocabulary, every scope scores near zero and the
winner is whoever took the 0.15 recency tiebreak.
The three evidence gates are OR'd because they fail on different question
shapes. A short question about a rare symbol has high evidence and a low
count; a long question about a familiar subsystem has the reverse. Requiring
all three abstains on both.
Enumeration inverts every other gate, so it gets its own selection rule.
Every gate above asks is there enough evidence for ONE scope, and a question
about something several projects share splits its evidence across them by
construction — so the more projects genuinely share a term, the less likely all
of them come back. Measured, which of my projects use Cloudflare workers or D1? returned one owner of two: 2.098 against 1.0867, where the 0.85 ratio
needed 1.783.
The ratio is the part that inverts. It measures distance from the top scope,
and the top scope of an enumeration is merely whichever owner says the term most
often. Set mode drops it and asks each scope the same routability question in
turn, against a floor discounted to 0.8 — because a member of a set legitimately
carries less evidence than a lone answer.
Enumeration is grammatical, like deixis, so it is detected the same way: a closed
class of markers, no threshold. It also outranks deixis. "where else does
this pattern appear?" points at its subject and asks across the corpus; the
deixis rule exists because a pointing question gives no way to pick one scope,
and an enumerative question is not picking one.
The frame nouns are stripped from the query, which is the deixis insight one
level down — projects is scaffolding, not vocabulary. It is not a cosmetic
step: in the development corpus projects is held by exactly two scopes, so it
scored as discriminative evidence for those two and dragged both into every
"which of my projects" answer. Stripping it moved negative-family abstention
from 66.7% to 88.9%.
loci ask "which of my projects use Cloudflare workers or D1?"
# ENUMERATED -> 3M1RY33T.github.io, urthreads
Scopes and groups
Groups are overlapping labels over a flat scope set. There is no tree anywhere,
and nothing is ever merged.
scope the groups it carries
─────────────────────────── ──────────────────────────────────────
delroy monorepo me delroy
delroy/glasses sub-scope me delroy
delroy/client sub-scope me delroy client:acme
acme-api repo me client:acme
vendorlib repo vendor:someorg
me, vendor:someorg read from git provenance by `scan` or `groups infer`
delroy computed by `scan` from the filesystem: a repository
that splits gives its own id to every sub-project
inside it, and to itself
client:acme asserted by hand, with `loci group add`
Five scopes, four groups. delroy/glasses is not stored insidedelroy; it
carries the label delroy exactly as it carries me. client:acme spans a
monorepo sub-project and an unrelated repository, which a hierarchy could not
express at all. The two sub-scopes exist only because that repository declared
them in .loci.json, or was scanned with --split; without either, delroy is
one scope and the containment group does not exist. A parent excludes its
sub-scopes' subtrees from its own collection, so no file is counted twice and
neither vocabulary is inflated with the other's tokens.
What a group does to a question is its mode, and the mode answers two
different questions depending on how the group was reached:
reached through your cwd named with --group X
──────────────── ───────────────────────────── ────────────────────────
explicit nothing confines to X's members
soft (default) every outside scope's confines to X's members
evidence base × 0.5
hard confines to members, and confines to X's members,
abstains when the best and abstains when the
answer is outside best answer is outside
The right-hand column is the counter-intuitive half, so it is measured rather
than asserted. On a three-scope fixture — alpha and beta in group team,
vend outside it — one question, four policies:
loci route "how does the gizmo parser emitter handle sprocket calibration"
unconfined -> vend, alpha
--group team explicit -> alpha
--group team soft -> alpha
--group team hard -> ABSTAIN (out_of_group); candidates: alpha (sprocket), beta (calibration)
All three modes dropped vend, the scope that won unconfined. --group is you
asserting the answer is in here; the mode decides only what happens when it is
not. The shortlist under hard is wider than the selection above it on purpose:
it is what the abstention refused to choose between, not what it would have
answered from — beta is on it because calibration is a term only beta
holds, and off the answer because that is all it has.
Reached through cwd instead, soft demotes and does not confine. The same
question asked from inside alpha with no --group at all returns
alpha, vend: vend scores 3.39 against alpha's 6.18 — far outside the
0.85 widening band — and comes back regardless, because it is the sole holder
of gizmo, parser and emitter, and the concentrated tier forces every
holder of a shared-or-rare term into the answer. Demoted is not excluded.
Commands
command
purpose
loci setup [dirs…]
scan, graph, index, embed and calibrate in one pass (-y, --no-embed, --split)
loci update [dirs…]
refresh graphs, index, vectors and calibration for what is registered (--no-scan, --force)
loci scan <dirs>
discover git repos and register them as scopes (--split)
loci add <path>
register one scope explicitly (--alias, --glob)
loci scopes
list what is registered (--group)
loci uses
which of your projects reference another, with citations (--refresh)
loci groups
list groups, their resolved mode, and how many are in each
loci groups infer
label every scope me or vendor:<org> from git provenance
loci group set <group> --mode
how much a group confines: explicit, soft, hard
loci group add|rm <scope> <group>
edit one scope's membership
loci graphs [scope…]
build missing structure graphs via graphify
loci index
build the routing index and episode store (--force)
loci embed
encode episode chunks locally (--model)
loci calibrate
fit the evidence floor to your corpus (--show)
loci route "…"
show where a question routes (--explain, --group)
loci ask "…"
route, then query both stores (--scope, --group, --fast, --rerank)
loci doctor
coverage gaps per scope, and the fix for each
loci eval
measure routing on your corpus (--misses)
loci skill install
install the /loci skill for an AI client (--dir, --print)
loci mcp
run the MCP server on stdio
loci index reuses any scope whose files and git HEAD are unchanged, so
reindexing is cheap enough to run from a commit hook.
Keeping it current
Nothing updates itself. There is no watcher, no daemon and no installed git
hook: every refresh is a command you run.
loci update # the whole chain, for everything registered
It is setup for an install that already exists, so it prompts for nothing and
opts you into nothing new: embeddings are re-encoded because you already had
embeddings, and the thresholds are refitted because they were already fitted.
It also rescans wherever your last scan was pointed, so a repository created
since then is registered rather than missed — that is what the roots key in
scopes.json is for, and loci update ~/work adds another.
The one thing it rebuilds unconditionally is the structure graph, because
that is the only staleness in loci that is otherwise silent. loci graphs
builds a graph that is missing and skips every scope that already has one, so
a project whose code moved keeps routing on the symbols it had the day it was
registered. Everything else announces itself: index re-parses a scope whose
fingerprint moved, and doctor names embeddings that no longer line up with
the store. Rebuilding all of them is affordable because graphify does not
rewrite a graph whose code did not change — measured, a second consecutive run
left graph.json byte-identical with its mtime untouched, and loci's own
fingerprint reads that mtime, so refreshing the graphs does not force a reindex
of the scopes that stood still.
By hand, the same thing, and the order is the dependency chain again:
loci graphs --all # `loci graphs` alone would skip every existing graph
loci index # unchanged scopes are reused, not re-parsed
loci embed # a full re-encode; skipping it silently disables
# semantic search for every scope whose chunk count moved
loci calibrate # the floor measures shared vocabulary, which just changed
Two things update cannot do for you. A project deleted from disk stays in the
registry — it is reported, not removed, because the entry carries groups and
aliases you set by hand and the directory may only be unmounted. And a
long-running loci mcp server holds its fitted rankers and vectors in memory
from boot, so restart it after a rebuild.
From inside an AI client
loci skill install # -> ~/.claude/skills/loci/SKILL.md
That installs one skill with subcommands, not eight commands cluttering the
listing:
/loci "<question>" ask memory; routes on your working directory
/loci update refresh everything registered
/loci doctor coverage gaps per project, and the fix for each
/loci scopes what is registered, and the group each is in
/loci route "<question>" where a question routes, and why it abstained
/loci add <path> register one project
/loci setup [dirs…] first run
/loci <anything else> passed through to the `loci` CLI verbatim
The passthrough rule is what keeps it short: index, embed, graphs,
calibrate, eval and groups all work from the client without being
documented twice, so the skill cannot drift from flags it never mentions. What
it does mention is checked — a test parses the Usage block and fails if any
command in it stops existing.
The skill ships inside the package rather than in a dotfiles repo, because it
documents flags and abstention reasons that the CLI can change; loci skill install after an upgrade is the whole update path. --print writes it to
stdout for a client that does not read ~/.claude/skills.
Two things it teaches an agent that are not obvious from the CLI. Run it from
the project root — cwd is the primary routing signal, so a question asked from
the wrong directory routes worse than one asked from none. And ABSTAINED is
an answer: the reason names what to do next, the shortlist under it names the
projects that could plausibly own the question and the terms that put them
there, and an agent that rephrases and retries instead is burning turns on a
decision that was about vocabulary rather than wording. An abstention carrying no
shortlist at all is the one case where rephrasing cannot help either — nothing in
the corpus holds a distinctive term from the question.
MCP
loci mcp speaks stdio, so a client launches it — you do not run it yourself.
Claude Code:
claude mcp add loci --scope user -- loci mcp
claude mcp list # loci: ... - ✔ Connected
Use an absolute path for command if loci lives in a version-managed
environment — pyenv shims and pipx venvs resolve against a shell your client
does not have. which loci from a working shell, then paste the target.
Three tools — ask, scopes, doctor — not several dozen. Agents are the
primary consumer, and a wide tool surface pushes orchestration onto the model
and burns a turn per hop.
Without this the server is dead weight: it is installed by the mcp extra and
reachable by nothing, and every question an agent asks goes through the CLI
instead — paying full process and model startup per question, and never touching
the boot-time warm-up the server exists to provide.
Pass cwd to ask whenever the client knows it, and group to narrow the same
way --group does on the command line. The server loads the embedding model and
the largest scopes' rankers at boot, so the first tool call an agent makes is
fast rather than paying ~4.4s of startup while the user waits.
Measure it on your own corpus
Every accuracy figure below was fitted and measured against one person's
repositories — ten when the constants were fitted, fourteen now. That is a claim
you would otherwise have to take on trust, and the drift between those two
numbers cost the behavior family 57 points.
loci eval # a few seconds, no labelling, no setup
loci eval --misses # and what it got wrong
It only asks questions whose correct answer is known by construction:
family
how gold is known
deictic + cwd
asked from inside a scope, so the answer is that scope
deictic, no cwd
nothing says which project — abstaining is the only correct answer
unanswerable
nonsense no software corpus can answer
signature
built from each scope's own most distinctive vocabulary
10 scopes | random guessing would score 10.0%
family n correct scopes what it measures
deictic + cwd 80 100.0% - should route to the scope you are in
deictic, no cwd 8 100.0% - should abstain; nothing says which project
unanswerable 6 100.0% - should abstain; no corpus can answer these
signature 20 100.0% 1.0 are your scopes distinguishable from each other?
The signature family is an upper bound — it asks each scope about its own
rarest words, so a real question phrased in shared vocabulary does worse. A low
score there is a fact about your projects, not necessarily a bug to tune away.
Calibration
Routing gates on three independent signals, and any one passing is enough:
one exceptionally discriminative token, enough summed token evidence, or enough
matched tokens. They fail on different question shapes — a short question about
a rare symbol has high evidence and a low count, a long question about a
familiar subsystem has the reverse — so requiring all three abstains on both.
The evidence floor is the one that depends on your corpus, because it depends on
how much vocabulary your projects share. loci calibrate fits it from the same
auto-labelled questions, and reports whether the bands separate at all:
evidence floor 7.17
fitted from 149 routable + 14 unroutable questions, none hand-labelled
classifies 84.0% of its own samples correctly
! The bands OVERLAP -- some questions that should route score no higher than
questions that should not, so no threshold separates them cleanly.
An overlap warning is information about your projects, not a defect: they share
enough vocabulary that evidence alone cannot always tell them apart, and cwd or
--scope should carry more of the weight.
What it reads, and what it writes
Everything stays on your machine. No LLM in the query path; routing is a dict
lookup per token. Embeddings, when enabled, run locally.
Collected per scope: README*, docs/**/*.md, source files (for docstrings and
comment blocks), and git log. Add anything else with loci add --glob,
including absolute paths outside the repo.
The only configuration loci reads from inside a repository is .loci.json, and
only if you write it. It names sub-projects at any depth, and unlike
--split it is honoured on every scan:
Paths are relative to the repository root and may not escape it. A malformed
file is ignored — a scan that aborts on a stray comma is worse than one that
misses a declaration. Nothing writes this file; you do.
The only thing loci ever puts inside a repository is graphify-out/graph.json.
Two commands write it: loci graphs, which does nothing else, and loci setup,
which offers it at step 2 of 5 and takes yes as the default — including with no
terminal, where every prompt takes its default. loci setup --no-graphs skips
it, and every other command reads.
So inside a repository exactly two files matter, and they point in opposite
directions:
<repo>/graphify-out/graph.json loci writes it, on request. You never do.
<repo>/.loci.json you write it. loci never does.
Everything else goes to ~/.loci (or $LOCI_HOME). One directory, not one per
project: routing is a cross-scope decision, so answering which project is this
about needs every scope's vocabulary in the same lookup, and N per-project
indexes would turn the cheap question into the expensive one.
~/.loci/ what it holds written by
├── scopes.json the scope registry, and scan, add, group,
│ the roots it was scanned update
│ from, for `loci update`
├── groups.json group policy: default mode group set
│ and any mode set per group
├── scope_index.json routing index, the token index
│ -> {scope: node_df} map
├── episodes.json the episode store: chunks index
│ verbatim, already redacted
├── rankers/<scope>.lex fitted BM25 + char-gram index
├── embeddings.npz chunk and symbol vectors embed
├── .symbols-<scope>.json symbol labels, in the embed
│ order embeddings.npz has
├── calibration.json the fitted evidence floor calibrate
│ + a semantic floor / scope
└── .index.lock advisory build locks, so index, embed
.embed.lock two builds cannot inter-
leave their writes
scopes.json is machine-managed and rewritten wholesale by every scan;
groups.json is yours and must survive one. That is why membership lives in the
registry and mode lives in the policy, and why they are two files.
Redaction
Everything collected passes through loci.redactbefore it is written, so a
credential committed by accident never reaches the episode store, the vectors or
the rankers.
Files that exist to hold secrets are never opened — .env*, .dev.vars,
*.pem, *.key, id_rsa, secrets.*, credentials.json. Then every chunk is
scanned for AWS keys, GitHub/Slack/Google/Stripe/OpenAI/Anthropic tokens,
private-key blocks, JWTs, user:password@host connection strings, bearer
tokens, and long opaque values assigned to a secret-shaped name.
Chunks are built through a single constructor, so a new collection path cannot
forget to redact. A secret assignment keeps its key name and loses only its
value, so the chunk still reads sensibly and still routes:
loci index reports what it removed, per scope and by kind — silent redaction
is indistinguishable from none. Deliberately biased toward over-redaction: a
false positive costs one chunk a little retrieval quality, a false negative
copies a live credential into a plaintext file and a vector index.
Design rules, and the measurements behind them
Each of these is something a reasonable implementation gets wrong.
1. Never merge scopes into one index. See the table at the top: 2% and 18%
on-topic against a merged graph.
2. Abstention is a feature at both layers. Measured 10/10 on probes
including deliberate nonsense. Every RAG tool confidently answers "what is the
airspeed velocity of an unladen swallow"; this one says it does not know.
3. Margin and ratio tests do not measure confidence. Both were tried and
both invert on real data. "How do I fix this bug?" produces routing margin
0.94 while a correctly-routed question produces 0.18. Nonsense scored 0.608
against a large scope with a higher top-to-p90 ratio than a genuine question
against a small one. What separates them is whether the question's words exist
in the corpus at all.
4. A matched-token count is not evidence. Compared against auto-labelled
questions, count does not separate routable from unroutable at all — it inverts.
Summed token evidence does. Count is kept only as one of three OR'd gates.
5. Deixis is grammatical, so detect it grammatically. A question that points
at its subject ("this project", "the app", "it") cannot be routed by vocabulary,
because the words that would identify the subject are exactly the ones it
declines to say. No lexical statistic rescues this: per-token evidence for
start (2.51) and services (2.56) is indistinguishable from real evidence
like session (2.53). A closed class of markers took abstention from 37.5% to
100%.
6. Normalize for scope size, twice. With one scope 67× larger, ordinary
English words appear in exactly one scope and score as maximally discriminative
for it.
7. Docstrings are excellent retrieval material and poor routing material.
They are high-volume generic English — "returns", "the value", "default",
"configuration" — and folding them into routing halved accuracy on both a
symbol-indexed and a prose-only corpus.
8. Coverage is the binding constraint, not ranking. This surfaced in every
experiment. Running a free AST-only index over three un-indexed repos moved
routing 69.6% → 82.6% while adding two more scopes to compete against. Hence
doctor as a first-class command: an empty answer with a reason beats a
confident answer from the wrong project.
Ranking
Episode search fuses three signals, renormalizing over whichever produced one:
ranker
catches
BM25
exact terminology, rare identifiers
char 3–5 grams
morphology and casing — samesite vs SameSite=None
embeddings (bge-small, local, optional)
meaning without shared words
The lexical pair recovers questions the structure graph cannot see at all. The
embedding ranker covers what neither lexical ranker can: a question and its
answer that share no vocabulary.
Adding embeddings destroyed abstention — every chunk gets a nonzero score, so an
absolute floor stops firing. The fix is a two-tier gate: a hit must be
lexically grounded or semantically confident, because a lexical gate alone
rejects exactly what embeddings were added for.
Cross-encoder reranking is available via --rerank and off by default: it
measured 2/6 → 3/6 precision@1 at ~96ms/query with two small regressions, on six
cases whose labels are themselves arguable.
Performance
loci index 15s full · 3.9s when nothing changed
loci ask 1.0s one-shot with --fast · ~5s with semantic ranking
MCP server 6s at boot, then 0.2-0.3s per tool call
routing sub-millisecond, and flat from 25 to 100 scopes
Three things make that work, each a measured bottleneck first:
Traversal prunes, it does not filter.Path.glob("**/*.py") descends into
node_modules, .venv and Pods in full — 32.9s to enumerate one repo. It
improves the index as well as the clock: one project was 66% vendored
third-party source before pruning.
Lexical rankers are fitted at index time. Char 3–5 gram TF-IDF over a large
scope costs ~1.5s, and an in-process cache never survives a CLI or MCP
invocation.
The embedding model is loaded at MCP boot. ~2.3s of imports plus model
construction is process startup, not work. --fast skips it for one-shot CLI
questions.
Scale
scopes
vocabulary overlap
routing
top-1
25–100
low
0.1–0.5ms
100%
25–100
moderate
0.3–0.6ms
90%
25
high
0.1ms
36% — answers, and is usually wrong
50–100
high
0.5ms
abstains 90% of the time
Scope count is not the limit; vocabulary overlap is. The failure gets safer as
scope count rises — more scopes sharing a term means less evidence for any one,
and the gate fires. A small corpus of highly similar projects is the
dangerous configuration.
Durability
Writes go through a temp file and os.replace, so a reader never sees a
half-written index. Path.write_text truncates before it fills: measured on an
8MB store, 11 torn reads in a few seconds of concurrent access, and 0 after
the change. An MCP server reading while loci index runs sat squarely in that
window.
loci index and loci embed hold an advisory lock, so two builds refuse to
interleave rather than producing an index and a store that disagree. Locks left
by a dead process are detected and broken.
The store is written before the index, because the index is what readers gate
on: a crash between the two leaves a stale index pointing into a store that is a
superset of it, never an index promising chunks that do not exist.
Nothing above backends/ knows which implementation is in use. Swapping
graphify for tree-sitter, SCIP or an LSP index means writing one adapter, not
editing the router. searchmust be able to return an empty list — a store
that always finds something cannot be trusted when it does.
Why it is not a graphify fork
loci depends on graphify through one adapter that reads its graph.json data
contract and shells out to graphify query --graph. Forking would mean
inheriting a 767KB extractor and ~18 tree-sitter grammars to change code that
never needed changing. The scoping gap that motivated loci — graphify tags
merged-graph nodes with a repo attribute and exposes no way to filter on it —
is solved above graphify by keeping one graph per scope. Both projects are
MIT, so forking stays available; it just is not necessary.
Status
Alpha, and the numbers deserve their caveats.
Measured on a 173-item set over 14 scopes. 112 of those items are generated from
a fixed taxonomy applied to every scope, so that portion is unbiased by
construction. The remaining 61 were hand-authored by the same person who built
the router; each carries a contamination field recording what the author had
seen, and uncontaminated results are always reported separately.
The eval corpus is not in this repository. The harness reads fourteen real
projects, six of them private or never pushed anywhere, and its records name
them at file:line — so evals/ and the evals/RESULTS.md cited throughout
these sources are kept locally and are not published here or in the sdist. What
ships instead is loci eval, which generates its own questions from whatever
corpus you have registered; that is the number worth trusting about your own
setup, and it is the one this README asks you to run.
Known limits:
The behavior family fell from 85.7% to 28.6% when the corpus grew from 10
scopes to 14, and the cause is not a threshold — sweeping SIZE_PRIOR
across 0.0–0.5 on the new corpus never beats the shipped 0.15. Two scopes now
hold 5,995 and 4,680 distinct tokens against urthreads' 427, and a scope
whose vocabulary approaches the corpus's own matches every question on
ordinary English. Five alternative scoring families were swept against it and
none beat the plain sum; the one strict improvement ships inert as
CORROBORATION_WEIGHT, because a single item on a single corpus is not
evidence. This is the largest open problem in the router.
The synthetic test bed cannot reproduce that failure: CorpusSpec.size_skew
varies file volume, not vocabulary breadth, so all thirteen shapes read 100%
at every value of the constant that would fix it. Closing that is a test-bed
change and it blocks the router change.
Two eval golds were found stale during this release, one of which was marking
a correct three-scope answer down as a precision failure. Corpus drift
turns right answers into recorded misses and nothing in the harness notices.
The structure store surfaces the answering symbol in 4 of 7 probes.
Large prose-only scopes over-attract; a scope with thousands of chunks and no
code graph absorbs questions belonging elsewhere.
CI now runs the suite on Linux, Windows and macOS across Python 3.10–3.13.
Standing it up found a third platform bug that reading had missed: Windows
refuses os.replace against a destination another handle holds open, so every
durable write failed there while both other platforms stayed green. Reading
finds bugs — it did not find that one. No user has run loci on Windows yet.
loci eval and loci calibrate share question families, so a perfect eval
score immediately after calibrating is partly circular.
The most useful thing anyone can do is run loci eval on a corpus that is not
mine and say what it reports.
Chưa có danh sách công cụ. Máy chủ từ xa sẽ có danh sách sau lần kiểm tra tình trạng đầu tiên.
Câu hỏi thường gặp
Câu hỏi
Làm thế nào để cài đặt Loci?
Sao chép cấu hình stdio cho Cursor, Claude Desktop, Claude Code hoặc VS Code trên trang này. Cấu hình dùng bản cài đặt của nhà phát hành, không phải proxy của PlugMCP.
Loci có cần khóa API không?
Bản ghi này không liệt kê biến môi trường bắt buộc nào.
Loci dùng giao thức truyền tải nào?
Loci chạy cục bộ qua stdio.
Loci được kiểm tra lần cuối khi nào?
Kiểm tra lần cuối: 2026-11-08. Scoped memory for coding agents: picks the right project, answers in scope, or abstains.
