Agenten stärken
Der eigene Befehl des Herausgebers. Kein PlugMCP-Endpunkt.
Host a copy with our keys and logs.
npx -y @memharness/[email protected]Claude Code hat keinen Installationslink für einen lokalen Befehl. Kopieren Sie den Befehl unten und führen Sie ihn in einem Terminal aus.
MEMHARNESS_DB=<value> claude mcp add io.github.las7-memharness -- npx -y @memharness/mcp@0.1.2In einem Terminal ausführen.
{ "mcpServers": { "io.github.las7-memharness": { "command": "npx", "args": [ "-y", "@memharness/[email protected]" ], "env": { "MEMHARNESS_DB": "<value>" } } }}claude_desktop_config.json
Remote-URL oder ein Befehl auf Ihrem Rechner? Worin sich die beiden Installationen unterscheiden
Auf PlugMCP ausführen
A new server you create. It is not this listing's official endpoint. Use the publisher install when you want their URL and their credentials.
StartenVom Herausgeber
A bi-temporal, provenance-carrying memory primitive for AI agents. One SQLite file. No LLM or network calls in the storage layer. Exposed to any agent via MCP.
Most agent memory is a bag of strings. memharness stores facts, and combines three semantics that incumbents tend to split apart:
Bi-temporal, provenance-carrying long-term memory for AI agents, in one SQLite file.
valid_fromvalid_totx_atThe storage layer is deterministic: no LLM, no network, no background daemon. It's plain SQLite, so you can open the file with any client.
memharness is not a magic accuracy upgrade, and it is honest about that. If your
agent's memory is small and static and comfortably fits the context window, a
CLAUDE.md file (or just stuffing the history into the prompt) is simpler, and
on short histories full context will match or beat any external memory system.
Reach for memharness when:
as_of), "what changed since Monday?" (diff), "why does it
believe this?" (why). These are queries a bag of strings cannot answer.Honest, and pointed at the thing memharness actually does differently: it is a deterministic, auditable storage layer rather than an extraction service.
Where the others win, plainly: mem0 and Zep do automatic fact extraction
from raw conversation, which memharness deliberately does not (the write path
stays model-free; a client or skill decides what is worth remembering). Plain
CLAUDE.md needs no install at all. memharness earns its place when you need the
temporal and provenance queries the others don't offer.
The default install is small (SQLite plus the MCP SDK); the embedding model is opt-in, see Hybrid recall.
Claude Code:
claude mcp add memharness -- npx -y @memharness/mcp
Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json)
and Cursor (~/.cursor/mcp.json) use the same JSON shape:
{
"mcpServers": {
"memharness": { "command": "npx", "args": ["-y", "@memharness/mcp"] }
}
}
Codex (~/.codex/config.toml) uses TOML, not JSON:
[mcp_servers.memharness]
command = "npx"
args = ["-y", "@memharness/mcp"]
The database lives at ~/.memharness/memory.db (override with MEMHARNESS_DB;
XDG_DATA_HOME is honored on Linux). Nothing else is written unless you turn on
the optional debug log.
remember.recall and answers from memory. Correct it and it calls revise;
the old belief becomes history, queryable with as_of / why / diff.No API key, no signup, no network. The first remember creates the SQLite file
and that's the whole setup. To watch the tools work end to end without an agent,
run the demo: cd examples && npm install && npm run demo.
By default the agent decides when to call recall. To push relevant memory in
at the start of every session instead (more reliable than hoping the model
remembers to look), add a Claude Code SessionStart hook that runs the
bundled memharness-context tool, whose stdout is injected into context:
{
"hooks": {
"SessionStart": [
{ "hooks": [ { "type": "command",
"command": "npx -y -p @memharness/mcp memharness-context --subject user" } ] }
]
}
}
It prints a compact dump of the most relevant current beliefs (and exits quietly
if there's nothing yet), so the agent starts each session already knowing the
durable facts. Pass --subject more than once to inject several entities.
import { Memharness } from "@memharness/core";
const mem = Memharness.open(); // ~/.memharness/memory.db
// Learn something now, then learn it was actually true earlier.
const { id } = mem.remember({
subject: "user",
fact: "lives in Osaka",
sourceRef: "session-2026-06-09",
});
mem.revise({ oldFactId: id, newFact: "lives in Tokyo", validFrom: "2026-05-01" });
mem.recall({ query: "lives" }).facts[0].fact; // "lives in Tokyo" (current belief)
mem.diff({ since: "2026-06-01" }); // { learned, revised, retracted }
mem.why(id); // { fact, ancestors, descendants }
recall returns a RecallResult ({ facts: ScoredFact[]; asOf; truncated; usedFallback }), not a bare string. asOf time-travels: mem.recall({ query: "lives", asOf: "2026-04-15" }) returns what was believed as held on that date.
That honors transaction time, so a fact learned today is not visible to a query
about the past.
Recall ranking is reciprocal-rank fusion over FTS5 BM25 (plus a vector rank when
hybrid recall is enabled), times confidence, times
recency decay (90-day half-life, configurable), scored in SQL. An optional
maxTokens budget caps output for context windows. A substring fallback catches
partial words and typos, in both FTS-only and hybrid modes.
By default, recall is FTS5 keyword search plus recency/confidence ranking: no model, fully offline. Hybrid recall adds a semantic leg via a local embedding model (BGE-small, ~130MB, downloaded once from the HuggingFace hub then fully offline: no API key, no per-query network). Enable it in two steps:
Install the optional embedding package alongside the server. With npx:
npx -y -p @memharness/mcp -p @memharness/embed memharness-mcp
(or npm i -g @memharness/embed for a global install).
Set MEMHARNESS_HYBRID=1 in the server's environment.
The server then keeps stored facts embedded automatically: facts you remember
become semantically searchable on the next recall, with no separate backfill
step. The first hybrid recall prints download progress to stderr while the model
loads. If the package isn't installed, the server says so and stays FTS-only; it
never fails closed.
At the library level, recall is embedding-provider-agnostic: pass your own query
vector to recall({ queryVector }) and attach document vectors with
setEmbedding(...), from any model you like.
Two sessions, weeks apart. The agent learns a preference, the user later corrects it, and a downstream question asks what the agent believed at the time:
// June 9: the agent learns a deploy target and acts on it.
const { id } = mem.remember({
subject: "project:acme",
fact: "deploys via Heroku",
sourceRef: "session-2026-06-09",
});
// June 16: turns out the team moved to Fly back on June 1.
mem.revise({
oldFactId: id,
newFact: "deploys via Fly.io",
validFrom: "2026-06-01",
sourceRef: "session-2026-06-16",
});
mem.recall({ subject: "project:acme" }).facts[0].fact; // "deploys via Fly.io"
// "Why did the CI config you wrote on June 9 target Heroku?"
mem.recall({ subject: "project:acme", asOf: "2026-06-09" }).facts[0].fact;
// "deploys via Heroku": what the agent honestly believed that day.
mem.why(id); // the full chain: Heroku, superseded by Fly.io, with sources.
mem.diff({ since: "2026-06-15" }); // surfaces the Heroku -> Fly.io revision.
No bag-of-strings memory can answer the as_of question, because it overwrote
Heroku the moment it learned Fly.io.
The property suite is the heart of the project: for randomized sequences of
remember/revise/forget, recall({asOf: T}) must equal the belief set produced
by a naive, SQL-free replay of the event log, probed at every event
timestamp ±1ms. 10,000 cases run on every push to main.
Benchmarked at 100k facts (10% revision chains, 2% retractions) on a developer
laptop (Apple Silicon): overall recall p95 ~1.3ms against a 10ms budget,
across four query shapes (two-term keyword, keyword + subject, subject-only, and
as_of + keyword). pnpm bench seeds the database and asserts the budget, so
the number is reproducible rather than quoted.
One deliberate divergence from the original prototype: retraction stores a
timestamp (retracted_at), not a flag, so as_of queries before the
retraction still see history, which is what the prototype's docs promised but
its SQL didn't deliver.
pnpm install
pnpm test # unit + behavior suites (property tests at 200 runs)
pnpm test:property # 10k randomized property cases
pnpm bench # seed 100k facts, assert recall p95 < 10ms
Schema migrations are forward-only, driven by PRAGMA user_version. Rows are
never deleted (forget tombstones), so facts.id doubles as the insert
sequence. All timestamps are canonical fixed-width UTC ISO 8601, making
lexicographic comparison chronological.
For debugging or measuring your own usage, set MEMHARNESS_DEBUG=1 and the
server appends an op-name and timestamp line (never fact content) to a
usage.log next to the database. It is off by default, fully local, and never
networked.
Apache-2.0
Bi-temporal, provenance-carrying long-term memory for AI agents, in one SQLite file.
{ "mcpServers": { "io.github.las7-memharness": { "command": "npx", "args": [ "-y", "@memharness/[email protected]" ], "env": { "MEMHARNESS_DB": "<value>" } } }}Tools
Noch keine Tool-Liste. Remote-Server erhalten eine nach dem ersten Health-Check.
FAQ
Kopieren Sie auf dieser Seite die stdio-Konfiguration für Cursor, Claude Desktop, Claude Code oder VS Code. Sie nutzt die Installation des Herausgebers, keinen PlugMCP-Proxy.
Der Herausgeber nennt diese Umgebungsvariablen: MEMHARNESS_DB.
Memharness läuft lokal über stdio.
Zuletzt geprüft: 2026-11-09. Bi-temporal, provenance-carrying long-term memory for AI agents, in one SQLite file.