Kb Recall

kb-recall

Persistent memory for Claude Code and GitHub Copilot, scoped to the feature you're working on and shareable with the rest of your team — and measured, because context is a budget.

Table of Contents

• 30-Second Quickstart Walkthrough
• The problem
• What makes this different
• Installation
• Troubleshooting
• Changelog
• Configuration
• KB structure
• How it works
• Tools
• Slash commands
• GitHub Copilot
• Version history
• Design notes
• Development
• Setup checklist
• Known problems
• Roadmap

30-Second Quickstart Walkthrough

Three commands, one reload, and your agent stops starting from scratch.

uv tool install kb-recall     # 1. install the CLI — isolated env, exposes one `recall` command
cd ~/your-project             # 2. setup is per-project — cd into the target first
recall setup -p claude        # 3. register MCP server + config + CLAUDE.local.md + hook
recall setup -p copilot       # 3b. (optional) Copilot: MCP config + hooks + Agent Skills

-p is required: setup writes into one client's configuration, so the command that picks which client is never inferred. -p copilot writes .mcp.json + .github/ and never touches ~/.claude/; -p claude is the mirror image. Run both to serve both.

Reload your client, then open a session on a feature branch. There is no load command to remember — the session hook reads your branch and loads the matching KB as the session opens. That is the moment in the GIF above. If the branch matches no known KB, your agent offers to create one (/recall:init) or link an existing one (/recall:link-feature).

Verify the install with recall --help. The full walkthrough, with a screenshot per step, is in GUIDE.md. For Copilot, see GitHub Copilot.

The problem

Your agent is capable within a session but forgets everything between sessions. Each new conversation starts from scratch: no knowledge of past bugs, no memory of architecture decisions, no awareness of the edge cases you already discovered.

Pasting context into every session is tedious and incomplete. And putting it in CLAUDE.md loads all of it, every session — one file per repo, speaking to everything at once, so a payment-gateway session pays for auth context it will never touch. Neither records anything either: whatever a session teaches you is gone unless you stop, notice, and write it down yourself.

What makes this different

Most memory layers ask one question: did it remember? This one asks four — was it even the right memory to load, did anyone have to write it down, who else gets to use it. Every mechanism here spends tokens, so every mechanism is measured.

• Feature-scoped, not project-scoped — and your branch picks for you. CLAUDE.md is one file per repo and it speaks to everything at once, so every session loads all of it. A KB is per feature, and the hook reads branch and loads the matching one as the session opens: a payment-gateway session never pays for the auth context. Switch branches and you switch KBs — there is no load command to remember.
• The KB is written while you work — you never take notes. Your agent calls save_memory in the turn it learns something, not at the end when the context is gone: a root cause, a constraint the code doesn't show, an approach tried and rejected, a decision with non-obvious reasoning. What lands is tagged, dated, and attributed to the journal it came from.
• Capture is automatic; promotion is gated. The two layers are built for opposite jobs. A memory is cheap and additive by construction — prepend-only, never edited, never deleted — so a write the agent makes on its own cannot damage the KB. The curated README.md is the opposite: a [gotcha], [rule] or [decision] reaches it only through a risk-classified promotion, and anything that would remove, merge or supersede an existing entry comes back as a diff to review before it is written.
• None of it is locked in a service — a KB is plain markdown. A feature's KB is a few .md files. Read it with cat, grep it from a script, diff it in a PR. No database, no daemon, no vendor account standing between your team and its own knowledge.
• One journal per contributor, so a team's findings accumulate instead of overwriting. Each engineer's agent writes to that engineer's own memories-{username}.md, while the shared README.md holds only what the team has agreed on. load_feature_context merges every contributor's journal present in the KB, chronologically, at read time — work one engineer did last week is in front of the next engineer this week — and because no two people write the same file, simultaneous saves never conflict.
• Shareable by git push, scoped so sharing stays safe. ~/.recall-mcp/<project>/ is a git repo from the first write, so handing a KB to your team is a push to a remote you control; and the repo is scoped per project, not per feature, so one push carries that project's KBs and never another project's. kb-recall only ever runs git init — committing stays yours.
• Cost guards, not just features. The UserPromptSubmit hook will block a prompt outright rather than let a session resume into an expired prompt cache with a large context — paying a full cache-write is worse than not running at all. See Idle-gap cost guard.

Installation

Requires uv — or pipx, which works identically. Two install sources are available — pick whichever your network allows:

From PyPI (recommended — stable releases):

uv tool install kb-recall
# or, with pipx:
#   pipx install kb-recall
cd ~/your-project
recall setup -p claude

From GitHub (latest main, or if PyPI is blocked on your network):

uv tool install git+https://github.com/nguyenquangsang/kb-recall.git

Either way, the CLI installs into its own isolated environment and exposes a single recall command (~/.local/bin/recall by default) — no manual git clone needed for the GitHub source either. This matters beyond convenience: the UserPromptSubmit hook runs recall prompt, and because the tool's environment is isolated and self-contained, that command reaches the right interpreter no matter which project's venv (if any) happens to be active in the calling shell — unlike invoking a bare python3 /path/to/prompt_submit.py, which only works when the ambient python3 happens to be kb-recall's own interpreter.

recall setup -p claude writes the hook using the absolute path to that executable rather than the bare name. Both harnesses (Claude Code and Copilot) spawn hook commands through a shell that does not source your interactive rc (.zshrc), so a bare recall — which resolves only because .zshrc puts ~/.local/bin on PATH — would match nothing there and the hook would fail silently, injecting no context and printing no error.

Verify:

recall --help

Contributing to kb-recall itself? Clone it and install in editable mode instead, so code edits take effect immediately without reinstalling:

git clone [email protected]:nguyenquangsang/kb-recall.git ~/kb-recall
cd ~/kb-recall
uv tool install --editable .

Run setup from your project directory (one-time per project):

cd ~/your-project
recall setup -p claude

This registers the MCP server, adds the project to config, writes CLAUDE.local.md (gitignored, per-developer — not the team-shared CLAUDE.md), and installs the hook. Reload your client after. See Setup checklist for what each step does.

Setup detects the target project from the current working directory (Path.cwd()) — always cd into the target project first.

Troubleshooting

Run recall doctor for a self-diagnostic — it prints a checklist and tells you what to fix:

$ recall doctor
✅ Python 3.12.1
✅ recall v0.3.0 @ ~/.local/bin/recall
✅ Claude Code @ /usr/local/bin/claude
✅ project registered (/path/to/your-project)
✅ KB initialized (2 feature(s))
✅ Claude UserPromptSubmit hook
✅ Copilot .mcp.json
✅ Copilot hooks (.github/hooks/recall.json)
✅ Copilot instructions (.github/copilot-instructions.md)

Every ❌ line carries a → hint naming the fix. The common ones:

• Python too old. kb-recall needs Python 3.10+. uv tool install fetches its own Python automatically, so a system Python 3.9 doesn't block it — this only matters for pipx, which uses your ambient Python: run pipx install --python python3.10 kb-recall, or upgrade Python first.
• recall: command not found. The install puts recall in ~/.local/bin (macOS/Linux) or %USERPROFILE%\.local\bin (Windows), which may not be on your PATH. Run uv tool update-shell and open a new terminal — or, for the current Windows session:

  $env:Path = "$env:USERPROFILE\.local\bin;$env:Path"
  

• No tools appear in the agent after a successful install. Two usual causes: the recall on your PATH is an older install than the one you upgraded (doctor prints the version it is actually running), or this project is missing its hook/config files (doctor lists each one and its fix).

Configuration

~/.recall-mcp/config.json is created automatically by setup. To add more projects manually:

{
  "projects": ["/home/you/my-project", "/home/you/another-project"]
}

Idle-gap cost guard

The UserPromptSubmit hook blocks a prompt outright (Claude is never invoked) when you've been idle long enough that the 1-hour prompt cache has likely expired and the session's context is already large — resuming would pay a full cache-write instead of a cheap cache-read. Lighter, non-blocking reminders fire earlier (idle alone, or context size alone) regardless of this setting. To disable just the hard block:

recall block-idle false   # disable
recall block-idle true    # re-enable (default)

This writes "block_idle": false to config.json. See GUIDE.md's "Idle-gap cost guard" section for the full tier breakdown.

Claude Code only. This guard works by blocking a turn, and Copilot's prompt-time event is mutation-only — it can rewrite what the model receives but cannot stop the turn. There is no Copilot equivalent; see GitHub Copilot.

KB structure

~/.recall-mcp/
├── config.json
├── usage.log                        ← every tool call, human-readable (tail -f friendly)
├── usage.jsonl                      ← same data as JSON Lines, for jq/pandas/LLM analysis
└── my-project/
    ├── .git/                        ← one repo per project, created on first write
    ├── features.md                  ← index of all features
    └── payment-gateway/
        ├── README.md                ← stable knowledge
        └── memories-{username}.md   ← prepend-only journal, one per contributor

How it works

Each feature gets a small knowledge base (KB) stored in ~/.recall-mcp/{project-name}/{slug}/:

An index file (features.md) lists all KBs for quick discovery.

The KB lives in ~/.recall-mcp/ — completely outside the project repo. It never gets committed accidentally, and you can version it independently by committing to that project's KB repo.

Your agent reads and writes these files via 8 MCP tools. No manual file management needed.

Every tool call is logged to two files: usage.log (human-readable, tail -f friendly) and usage.jsonl (JSON Lines for analysis with jq, pandas, or an LLM). ~/.recall-mcp/<project-name>/ is initialized as a git repo automatically on first write — git init only, kb-recall itself never auto-commits. That scope is deliberate: one repo per project, not one for all of ~/.recall-mcp/, so sharing one project's KB with its team never carries another project's KB along with it. This leaves it ready for you to commit/push manually if you want to version or share the KB.

Tools

list_features

Lists all feature KBs across configured projects. Shows names, slugs, summaries, and last-updated dates. Call this at the start of a session to orient yourself.

list_features(project="")

search_features

Keyword search across every feature's README.md and memories-*.md (all contributors) in scope. Space-separated keywords are OR-matched, case-insensitive. Returns lightweight snippets grouped by slug — not full content. Use this to check a specific fact that might live in a feature you aren't currently working in (e.g. "was X already done elsewhere"); not for slug discovery, which list_features already covers. Pass a comma-separated list ("proj-a,proj-b") to search a specific subset of projects, or "all" for every configured project — only when the user explicitly asks for a cross-project search; default stays scoped to the current project in every other case. The agent should not decide subset-vs-all on its own either — ask the user, offering the real configured project names as choices. Multi-project results are prefixed <project>/<slug>; the footer instructs the agent to show that list to the user and let them pick which KB to load, rather than deciding on its own.

search_features(query="webhook retry idempotent", project="")
search_features(query="webhook retry idempotent", project="proj-a,proj-b")
search_features(query="webhook retry idempotent", project="all")

load_feature_context

Loads the full README.md + memories-{username}.md for a specific feature. Use this before working on a feature to restore context from past sessions. If the feature's README has a related_tickets section, a hint listing related features is appended to the response. The response header always shows <project>/<slug> so it's unambiguous which project's KB was loaded, even when the same slug exists in more than one project.

load_feature_context(slug="payment-gateway", project="")

save_memory

Prepends a single engineering insight to the current user's memories-{username}.md, dated today. Only call this for significant findings: bugs, edge cases, trade-off decisions, breaking changes. Skip routine implementation details. Promotion to README is automatic at save time: [gotcha]/[constraint] → critical_warnings, [decision] → architecture, [rule] → business_rules; [bug]/[idea]/[pattern] stay in memories only.

save_memory(slug="payment-gateway", content="Stripe webhooks can arrive out of order — always reconcile against DB state, not event order.")

init_feature

Creates a new feature KB directory from templates in templates/, and adds a row to features.md (creates features.md if it doesn't exist yet).

init_feature(name="Payment Gateway", slug="payment-gateway", summary="Handles Stripe checkout and webhook reconciliation", project="my-project", ticket="PROJ-1234", branch="feat/payment-gateway")

project is required when multiple projects are configured. ticket is optional.

update_readme

Replaces or appends the content of a named XML section in a feature's README.md. Call this after init_feature to fill in sections, or whenever architecture or business rules change. Valid sections: overview, key_files, technical_stack, business_rules, architecture, critical_warnings, open_items, checklist, related_tickets.

mode="append" writes immediately. mode="replace" (default) is two-step: the first call (confirm=False, default) writes nothing and returns a real diff of old vs proposed content; call again with confirm=True to actually write.

update_readme(slug="payment-gateway", section="business_rules", content="**[rule] ...**")   # returns a diff, writes nothing
update_readme(slug="payment-gateway", section="business_rules", content="**[rule] ...**", confirm=True)   # writes for real

update_feature_index

Updates one cell (ticket, branch, or summary) and the Last Updated date on a feature's row in features.md — the index row itself, not README.md content. Called by /recall:save (Step 8) when a feature's actual scope has drifted from the summary set at init_feature, and by /recall:link-feature when linking a renamed branch to an existing KB; not meant to be called mid-task otherwise.

Same diff-then-write pattern as update_readme: confirm=False (default) previews a diff, confirm=True writes it. append=True adds , {value} to the existing cell instead of replacing it (e.g. a KB now tracked by two branches).

update_feature_index(slug="payment-gateway", field="summary", value="Stripe checkout + webhook reconciliation + refund flow")   # returns a diff, writes nothing
update_feature_index(slug="payment-gateway", field="summary", value="Stripe checkout + webhook reconciliation + refund flow", confirm=True)   # writes for real
update_feature_index(slug="payment-gateway", field="branch", value="feat/payment-v2", append=True, confirm=True)   # keeps old branch, adds new one

report_miss

Records a context miss — a mistake the agent made that the KB should have prevented. Call this when the user points out that the agent ignored or lacked knowledge that should have been in the KB. Saves a [MISS] entry to memories-{username}.md so the gap is visible and can be fixed. Immediately after, call update_readme on the section the miss revealed as missing — do not stop at recording the miss.

report_miss(slug="payment-gateway", description="Forgot that webhooks can arrive out of order — this was already in README but the agent didn't apply it.")

Slash commands

Eight commands that wrap the most common workflows. Slug argument is optional for load, save, and miss — auto-detects from git branch, with fuzzy matching when provided.

/recall:load [slug]      ← /recall:load or /recall:load pay (matches "payment-gateway")
/recall:save [slug]      ← auto-detect, confirms before saving
/recall:list [project]
/recall:init [name]
/recall:miss [slug]
/recall:link-feature     ← link current branch to an existing KB (after branch rename)
/recall:compact [slug]   ← retire noise from memories (translate, compress, hide already-in-README entries)
/recall:tidy [slug]      ← rewrite README sections to be compact, English-only, high signal-to-token

Install (one-time, run from project directory):

recall sync-commands

Reload Claude Code. If you're working from a clone of this repo (editable install), recall sync-commands symlinks commands/*.md into ~/.claude/commands/recall/ — edit the source files and changes take effect immediately (no reload needed). Falls back to copying if your filesystem doesn't support symlinks; re-run recall sync-commands after editing in that case.

On Copilot the same eight exist as Agent Skills, invoked without the colon — /recall-load, /recall-save, /recall-list, /recall-init, /recall-miss, /recall-link-feature, /recall-compact, /recall-tidy — and the argument is passed after the command rather than interpolated: /recall-save webhook ordering note. Installed by recall setup -p copilot; see GitHub Copilot.

Version history

Git is initialized automatically in ~/.recall-mcp/<project-name>/ on the first write operation (git init only — kb-recall never auto-commits). Commit manually whenever you want a snapshot:

git -C ~/.recall-mcp/<project-name> add -A
git -C ~/.recall-mcp/<project-name> commit -m "KB snapshot"
git -C ~/.recall-mcp/<project-name> log --oneline
git -C ~/.recall-mcp/<project-name> diff HEAD~1

Design notes

Tool descriptions have a hard cap

Claude Code silently truncates a deferred MCP tool's description at roughly 2,000–2,150 characters. The cut is proportional, not fixed — a longer docstring loses a larger fraction of itself, not a fixed amount. This is a behavioral measurement taken against a live session, not something the client exposes, so it establishes nothing about why the cut lands there or whether it holds across Claude Code versions.

That finding is why Args: sits before OUTPUT in every tool on this server, and why CLAUDE.md carries a length budget. The two docstrings that produced it were both trimmed afterwards, so the numbers are documented rather than restated here as live figures: docs/toolsearch-truncation.md carries the measurements, the method, and the transcript route for checking it yourself — ~/.claude/projects/*/*.jsonl records what the model was actually served, so a rendered description can be compared against the docstring on disk.

Development

uv sync                 # install deps
uv run recall-server    # run locally (stdio mode)

Feature KB templates live in kb_recall/templates/ — edit them to change what init_feature generates. Eight files: feature-README.md and feature-memories.md (the per-feature KB), features-index.md (the project index), claude-md-snippet.md (the CLAUDE.md section shown when a project isn't configured), claude-command.md (the standard every commands/*.md file follows), copilot-skill.md (the standard every skills/*/SKILL.md file follows), copilot-instructions.md (the instructions file recall setup -p copilot writes), and claude-settings.json.

Tests and lint:

uv run pytest        # 466 tests: hook helpers, CLI, server tools, prompt_submit,
                     # commands + skills contracts, Copilot adapter, smoke scorer,
                     # docstring contract, end-to-end lifecycle, stdio smoke,
                     # text-encoding guard
uv run ruff check kb_recall tests scripts

Repo-local agents live in .github/agents/ — VS Code loads every .md file in that folder as a custom agent. kb-audit audits this project's own KB for stale or unverifiable entries; its tool list is an allowlist of read tools plus the three read-only MCP tools, so it reports findings but cannot write to the KB.

Setup checklist

Several things must be in place for your agent to reliably use kb-recall. recall setup -p claude automates all of them — run it once per project.

cd ~/your-project
recall setup -p claude

What setup does

Step 1 — Register MCP server — runs claude mcp add --scope user recall .... Verify with claude mcp list.

Step 2 — Configure project — adds the project path to ~/.recall-mcp/config.json and appends the kb-recall rules to CLAUDE.local.md (created if missing, added to .gitignore) — not the team-shared CLAUDE.md, since these are per-developer opt-in instructions. These rules are what make Claude actually use the tools — without them, Claude has no instruction to load KB before working or save insights after.

Step 3 — Configure issue tracker — asks once whether the project uses Jira or none/other, saved to config.json (issue_trackers.<project-path>) so /recall:init never has to ask again. Skipped (not guessed) when run non-interactively — /recall:init falls back to asking itself if this is still unset.

Step 4 — Install hooks — injects into ~/.claude/settings.json:

UserPromptSubmit (the recall prompt command, written as an absolute path) — runs on every prompt:

1. CLAUDE.local.md warning if the kb-recall section is missing (checks legacy CLAUDE.md too, for pre-existing installs)
2. Feature index for the current project — first prompt of session only, so Claude knows which KBs exist without paying index token cost on every turn
3. KB suggestion from branch name — once per session: if branch matches a known feature slug, Claude loads the KB; if no KB exists, offers to create one or link to an existing KB (/recall:link-feature)
4. Turn counter reminders — two independent cadences, both active only when a KB is loaded: every 8 turns, reminds Claude to call save_memory if nothing has been saved recently; every 16 turns, reminds it to call report_miss. Misses are much rarer than saves, so checking at the same rate would be mostly noise

This is the only hook kb-recall installs. (An earlier design also used a PreCompact hook for a pre-compaction save reminder; dropped in favor of the turn-counter reminder above, which covers the same need without the extra hook.)

Step 5 — Sync slash commands — symlinks commands/*.md to ~/.claude/commands/recall/.

Reload Claude Code after setup.

recall setup -p copilot runs only steps 2, 3 and 5b (its own MCP config, hooks and Agent Skills) and never touches ~/.claude/, and recall setup -p claude writes no Copilot files — neither invocation configures the other client. Copilot details: GitHub Copilot.

GitHub Copilot

kb-recall runs on GitHub Copilot in VS Code alongside Claude Code. Both clients share the same KB store (~/.recall-mcp/), so a project's knowledge serves whichever agent you happen to be using — no second setup, no second copy to keep in sync.

cd ~/your-project
recall setup -p copilot

Reload VS Code and start a new chat session — the auto-load hook runs at session start, so a session that is already open will not see it.

The two machine-specific files are gitignored because they hold absolute paths — each developer runs recall setup -p copilot once, exactly as the Claude Code setup above runs recall setup -p claude.

Enable .mcp.json first. VS Code Stable defaults chat.mcp.workspaceRootConfig.enabled to false, which makes it ignore .mcp.json silently — the agent has no tools and reports no error. Turn it on in Settings, or add the server through the MCP UI.

1. No idle-gap cost guard. The Claude-side guard works by blocking a turn; Copilot's prompt-time event is mutation-only, so there is nothing to block with. Per-turn reminders still exist, but they come from the post-tool-use hook instead.

2. No reload when you switch branches mid-session. Auto-load fires once, at session start. Switch to another feature branch in the same session and the previous KB stays in context — ask for the new one explicitly (/recall-load). This is the one case where "there is no load command to remember" does not hold.

3. Auto-load is partial, not the whole KB. The injected context is capped at 10 KB. A larger KB is truncated at a section boundary and carries a line telling the agent to call load_feature_context for the rest. Sections render in priority order, so rules arrive before the journal — but for a big KB expect one follow-up tool call before the agent has the full picture.

Per-turn reminders count tool calls, not turns. Copilot's post-tool-use payload has no turn identifier, so the cadence is measured in matched tool calls (every 16). A turn that only reasons and answers — calling no tools — gets no reminder, and the counter resets each session.

Tested on VS Code 1.139.1 / copilot-agent 0.67.0, 2026-09-25 → 2026-09-26:

• sessionStart fires and its additionalContext is honored — the model quoted an injected codeword back verbatim, and the injected context persists across later turns, not just the first.
• postToolUse fires on real sessions (~2 state entries per tool call), so the reminder channel is live.
• The MCP tools are reachable from a Copilot session, and .mcp.json parses in both wrapped and bare forms.
• The generated hook config is accepted — an unknown event key is skipped per entry, silently, without invalidating the file.

Documented, not inferred: the Local harness parses matcher for Claude compatibility and then ignores its value, so the hook runs on every tool call. That is consistent with the ~2 state entries per tool call measured here. The effect is bounded — counting more calls means reminders arrive earlier, never later, never silently absent.

Skill names cannot carry a namespace. /recall:load is not a style choice we skipped: the format allows only lowercase letters, digits and hyphens in a skill's name, which must also equal its directory name, and an invalid name makes the skill silently fail to load. The colon form becomes available only when skills ship inside an agent plugin — the harness then prefixes the plugin name (/my-plugin:test-runner) on its own. So kb-recall's /recall-load is a hand-rolled stand-in for a prefix the format will not let us declare.

One dead end, recorded so nobody re-tests it: userPromptTransformed never fires on VS Code Local. It is a Copilot CLI event, and VS Code's own prompt event has no context-injection field — "prompt-time injection on VS Code" is not a missing feature, there is no such channel. Live reference: VS Code hooks reference.

Known problems

Current (unfixed)

Reliability — the agent has no guarantee it will call load_feature_context before starting work. The hook injects the feature index, and CLAUDE.md rules instruct the behavior — but both are advisory. The agent can skip them. There is no hard enforcement mechanism at the protocol level.

No session-end checkpoint — (partially addressed) A turn counter reminder fires every 8 turns, prompting the agent to call save_memory (and a separate one every 16 turns for report_miss). The session can still finish without saving if the agent ignores these signals. On Copilot the reminder is keyed to tool calls rather than turns, so a turn that calls no tools gets no reminder at all.

Future (when KB grows or work spans features)

Cross-feature blind spots — each feature KB is isolated. When work touches two features, the agent must know in advance to load both. Two hint-on-load mechanisms exist: a curated <related_tickets> section, and key_files overlap mined live from every KB's <key_files> section, which surfaces "Possibly related (shared key_files, unconfirmed)" split into strong (ordinary overlap, or an exact shared path::symbol) and weak (the path is a hub, referenced by many features). Both are advisory — the agent can ignore them, and neither fires unless a KB is already being loaded.

KB staleness — memories-{username}.md entries are prepend-only and never expire. Over time, old entries about bugs that were fixed or decisions that were reversed accumulate. Retirement is manual: [supersedes:XXXX] and [resolved:XXXX] do drop an entry from the merged view, but only once someone knows its id and writes the tag — nothing detects that an entry has gone stale. The one automated exception is path-level: key_files references that no longer exist on disk are flagged on load. /recall:compact and /recall:tidy are the human-triggered cleanup path.

Single-project assumption — the hook and most tooling implicitly assume one active project. Multi-project setups work technically but the UX (specifying project= on every call, hook injecting all projects' indexes) becomes awkward.

Roadmap

Audited against the source on 2026-10-06. Checkboxes describe what ships today, not what was planned — if a claim here disagrees with the code, the code wins.

🚀 DX & Slash Commands (Short-term)

• CLI Setup: recall setup -p <platform> automates the installation steps — MCP server registration, project config, issue-tracker preference, hook install, slash-command sync.
• Cross-project resolution: Auto-detect and resolve ambiguous feature slugs, disambiguating by project when a slug exists in more than one, with a fuzzy (difflib, ≥0.8) fallback when an exact match isn't found.
• /recall:init auto-slugify: Derives the slug from the branch segment, pre-fills the ticket ID from a key-shaped pattern in the branch name, and defaults the username from git config user.name. Every command takes an optional slug and auto-detects from the branch otherwise.
• Single-section update command: Rewrite one named README section without the model composing the replacement text first. update_readme already writes a section, but it takes the finished content as an argument — there is no command that takes "section X, change it this way". /recall:tidy rewrites sections but in bulk and for compaction, not for a targeted edit.
• Multi-slug context loading: load_feature_context(slugs=["payment", "fraud"]) for cross-feature tasks. Today slug is a single string.

🧠 Context Engine & Auto-Discovery

• Smart linkages: <related_tickets> section hints related context on load.
• Cross-feature discovery by key_files: Every KB's <key_files> paths are mined into a live path→slug index, so loading one feature surfaces "Possibly related (shared key_files, unconfirmed)" for the others.
• File-path mapping (hook): Auto-detect the KB from the edited file path (src/.../cache.py) instead of branch name alone. The key_files index above runs inside load_feature_context — prompt_submit.py still detects on branch name only, so nothing proposes a KB from the file you just opened.
• Summary-injected hooks: The feature index injected on the first prompt of a session carries each feature's one-line summary, so the agent can weigh which KB matters without a discovery call.
• Semantic Search: Combine text search with local embeddings for large KBs.

🧹 Memory Lifecycle & Quality Control

• Session nudges: Two turn-counter reminders via UserPromptSubmit — save_memory every 8 turns, report_miss every 16. A Stop hook for session-end save was rejected for token cost; an earlier PreCompact save reminder was built and later dropped in favor of the turn counter.
• /recall:compact: Compress older journal entries to prevent unbounded context growth — translate non-English entries, condense verbose ones, and hide entries already captured in a README section. Surfaces itself through a size hint on load and after every write.
• Batch Curation Agent: Weekly scheduled workflow to audit KBs, promote repeated patterns to README, and surface stale/contradicted entries. The .github/agents/kb-audit workspace agent (see above) covers the audit half on demand, but it is manual and VS Code-only — nothing runs it on a schedule.
• KB health signals: The hook reports a branch-matched KB's own counts (N decisions · N warnings · N past mistakes) on the prompt that matches it, and every load_feature_context call flags key_files paths that no longer exist on disk; oversized memories or README sections trigger a /recall:compact or /recall:tidy hint after a load or a write.
• Health ratio: report_miss vs save_memory as a proportion, so an unhealthy KB is visible without reading the numbers. Counts are surfaced today; the judgment isn't.

👥 Team Collaboration & Ecosystem

• Auto Git Remote Sync: Background git pull/push on load/save — keeping team READMEs synced while keeping personal memories-{user}.md conflict-free.
• Provenance tracking: Add source field (PR, Jira ticket, manual) to memories and misses.
• GitHub Copilot support: Agent Skills (all 8 skills), sessionStart auto-load and postToolUse per-turn reminders, with .mcp.json and .github/hooks/recall.json generated by recall setup -p copilot. Three Claude-only gaps remain by design and are documented in GitHub Copilot: the idle-gap guard, mid-session branch reload, and prompt-time injection.