Project Memory MCP Server

project-memory MCP server

A small, local MCP server that gives AI agents (Claude Code, Cursor, …) a shared, persistent memory of the projects in a code folder — what each project is, decisions made, and every bug/issue faced during development.

It is stateless: every tool reads/writes plain files on disk, so multiple clients (and multiple machines) share one source of truth.

The model

Design rule: durable, low-volume facts go in AGENTS.md (auto-loaded). High-volume history (bugs) goes in issues.jsonl (queried via search_issues). This keeps the always-loaded context small while keeping everything searchable.

Works even where MCP is locked down

Some orgs disable third-party MCP servers via policy (e.g. GitHub Copilot's MCP allowlist enforcement). Because the memory is plain files, not a service, the core value survives that:

• The memory itself is just files. AGENTS.md is auto-loaded by the editor reading it — no MCP call involved — so a project's identity, decisions, learnings, and preferences still land in the agent's context.
• The policy is Copilot-scoped and per-client. It doesn't affect the same server in Claude Code or Cursor, and orgs running allowlist / registry-only mode can permit it — this server is published to the official MCP Registry (io.github.kaaustubh/project-memory-mcp).

Only the interactive tools (log_issue, search_issues, …) go over the MCP channel; the file-based memory keeps working without it.

Tools

• list_projects, get_project, search_memory — read project memory
• append_decision, append_learning — append a dated bullet to AGENTS.md
• remember_preference — turn a correction / stated habit into a remembered pattern (## Preferences in the root AGENTS.md for a global habit, or a project's for a local one); rides the auto-load, so it comes back next session
• log_issue — record a bug/problem → issues.jsonl
• search_issues — "have we hit this before?" across all projects (field-scoped; optional tags filter)
• list_open_issues, resolve_issue — track / close bugs
• sync_registry — reconcile the root AGENTS.md projects table with what's on disk (adds rows for new projects, flags stale ones)
• find_by_file — given a file path, surface the issues + decisions/learnings that touch it ("why is this code like this?")

You don't call these directly — you talk to your agent in natural language and it picks the tool. See Using it day to day below for what to actually say.

Using it day to day

Most of it runs itself: opening a project auto-loads its AGENTS.md (the agent already knows the project), and capture is proactive (plus the optional Stop hook). Your job is mainly to pull memory at the right moments. Just talk to your agent:

The one habit that matters: make "have we hit this before?" reflexive before every debugging session. That's where a memory tool earns its keep; the rest the system handles.

Capture is confirming, not silent — when the agent logs something it tells you in one line. Correct it freely: "don't log that", or "actually, log this too."

Escape hatches: PROJECT_MEMORY_HOOK=off silences the Stop hook for one session; uninstall-hook removes it entirely.

Install (npm — recommended)

From your code/projects folder, run:

cd ~/code            # the folder that holds your projects
npx -y @kaaustubh/project-memory-mcp install

That registers the server with Claude Code (user scope) and Cursor, using the current directory as your projects root. Restart those apps, then ask your agent "set up project memory for this folder" to scaffold AGENTS.md for each project.

No clone, no global install — the MCP config just runs npx, which fetches and runs the latest version on demand.

From source instead

git clone https://github.com/kaaustubh/project-memory-mcp.git ~/code/.memory-server
cd ~/code/.memory-server && ./install.sh

How it works (after install)

A common question: "once I install it, does it just start doing things?" Not quite — the server is passive. Here's the actual flow:

1. Restart your editor. MCP servers are loaded at startup, so the server only becomes available the next time you launch Claude Code / Cursor.
2. Push layer (automatic, not the server): when you open a project, the editor reads AGENTS.md (via CLAUDE.md → @AGENTS.md) into the model's context for you. This is why the agent "just knows" what your project is — it's a built-in editor feature.
3. Pull layer (the server, on request): the server announces its tools and then waits. It does nothing on its own. The agent calls a tool only when it's relevant — e.g. you say "log this bug" or "have we hit this before?", or the model decides a tool is useful. There's no background process or scanning.

Day one is empty. A fresh setup has no AGENTS.md files yet, so the auto-load has nothing to load and log_issue will refuse until a project's memory exists. Bootstrap once by asking your agent: "set up project memory for this folder" — it creates the AGENTS.md files. After that, everything works.

In short: a convention (auto-loaded files) + a tool the agent chooses to use + a one-time setup. No magic, no daemon.

Proactive capture (you don't have to say "log this")

The server ships a standing capture policy (sent to the client on connect, plus directive tool descriptions), so the agent records things on its own instead of waiting for you to ask:

• Before debugging a reported error → it checks search_issues for a prior fix.
• After fixing a non-trivial bug → it calls log_issue.
• After a real decision or a durable gotcha → append_decision / append_learning.
• After you correct how it works or state a habit → remember_preference, so the one-time correction becomes a pattern it brings back next session.

It's proactive but not silent: the agent tells you in one line what it recorded, asks when unsure rather than logging noise, and skips trivia and secrets. You can always override — "log this", or "don't bother". The standing policy is best-effort (it depends on the model following it); for a hard guarantee, add the opt-in Stop hook below.

Guaranteed capture (opt-in Stop hook)

The standing policy can be forgotten mid-session. The Stop hook makes capture non-optional: when the agent tries to end a turn, it runs once and blocks the stop to ask for one capture pass when either (a) real work happened (file edits or a commit) and nothing was written to project memory, or (b) you corrected how it works and no preference was saved. If memory was already written, or nothing changed and you didn't correct it, the hook stays silent and lets the turn end.

npx -y @kaaustubh/project-memory-mcp install-hook    # turn it on (then restart Claude Code)
npx -y @kaaustubh/project-memory-mcp uninstall-hook  # turn it off

• Off by default — plain install does not add it; you enable it explicitly.
• No loops — it fires at most once per turn (guarded by stop_hook_active), then lets the agent stop.
• Per-session kill switch — set PROJECT_MEMORY_HOOK=off to disable without uninstalling.
• Cost — it adds one extra model turn only on sessions that changed code but logged nothing, or where you corrected the agent and no preference was saved; silent otherwise.

Automatic recall (opt-in UserPromptSubmit hook)

Capture is only half the loop — the other half is remembering to look. The recall hook closes it: every time you submit a prompt, it matches your request against your issue history and decisions/learnings/preferences, and silently injects the strongest hits as context. So a prior fix or decision surfaces without you (or the agent) remembering to search — the "have we hit this before?" habit becomes automatic.

npx -y @kaaustubh/project-memory-mcp install-recall    # turn it on (then restart Claude Code)
npx -y @kaaustubh/project-memory-mcp uninstall-recall   # turn it off

• Semantic matching (when available) — if the optional embeddings model (@xenova/transformers) is installed, recall matches by meaning, so "the build is broken" still surfaces an issue logged as "compile failure" even with no shared words. Runs fully offline (the model is fetched once, then cached). Without it, recall falls back to keyword matching automatically — no configuration, nothing breaks.
• Silent unless relevant — injects nothing for trivial prompts or when there's no match.
• Ranked & capped — current-project hits rank highest; at most 4 lines are injected.
• Off by default — like the Stop hook, it's opt-in (per-prompt cost). Plain install adds neither hook.
• Per-session kill switch — set PROJECT_MEMORY_RECALL=off to disable without uninstalling.

Warm the cache: after a big logging session (or once, after enabling recall) run npx -y @kaaustubh/project-memory-mcp reindex to pre-embed everything, so the first recall isn't the one that pays for it. Vectors are cached per project in a derived .embeddings.json (safe to delete / git-ignore — the .jsonl + AGENTS.md stay the source of truth).

Pair it with the Stop hook and the loop runs itself: the Stop hook guarantees things get saved, the recall hook guarantees they come back at the right moment.

Across machines

The tool and your memory content sync separately:

1. Tool: nothing to sync — npx always pulls the published version (or git pull if you installed from source).
2. Content: each project's AGENTS.md + issues.jsonl live inside that project's own git repo, so cloning your projects brings their memory along. Nothing to copy.

issues.jsonl holds real bug details — only commit it into private repos.

New-project scaffold

For a new project under the root, create <project>/CLAUDE.md containing @AGENTS.md and a <project>/AGENTS.md with ## What this is, ## Stack & layout, ## Run / build / test, ## Decisions, ## Learnings sections.

Changelog

v1.6.1

• Docs: added "Works even where MCP is locked down" — clarifies that the file-based memory (AGENTS.md auto-load) keeps working even where an org disables third-party MCP servers (e.g. GitHub Copilot's MCP allowlist), since only the interactive tools use the MCP channel.

v1.6.0

• Semantic recall (optional local embeddings). The recall hook now matches your prompt against memory by meaning, not shared substrings — "the build is broken" surfaces an issue logged as "compile failure". Powered by a local, offline embedding model (Xenova/all-MiniLM-L6-v2 via the optional @xenova/transformers dependency); vectors are cached per project in a derived .embeddings.json, keyed by content hash so edited/removed items self-invalidate. If the model isn't installed it falls back to the previous keyword matching automatically — nothing to configure, nothing breaks. New reindex subcommand pre-embeds all memory so the first recall isn't slow. This completes the long-deferred "semantic retrieval" lever behind both recall and search_issues; keyword remains the zero-dependency floor.

v1.5.0

• Automatic recall (opt-in UserPromptSubmit hook). New install-recall / uninstall-recall subcommands register a hook that keyword-matches every prompt against your issue history and decisions/learnings/preferences and silently injects the strongest hits as context — so prior fixes and decisions surface without anyone remembering to search. Closes the other half of the capture↔recall loop. Silent on trivial/no-match prompts (generic filler words ignored), current-project hits ranked highest, at most 4 lines injected. Off by default; per-session kill switch PROJECT_MEMORY_RECALL=off.

v1.4.1

• Packaging: add the mcpName field (io.github.kaaustubh/project-memory-mcp) required to list the server in the official MCP Registry. No functional change.

v1.4.0

• remember_preference — corrections become remembered patterns. New tool that writes a dated bullet under ## Preferences, either in the root AGENTS.md (scope global — applies to every project) or a single project's. Because preferences live in the auto-loaded AGENTS.md, recall is free: a one-time correction ("never add a co-author trailer", "always typecheck before committing") comes back next session and is applied instead of re-corrected. Closes the cross-session loop for how you like to work, not just project facts.
• Correction-aware Stop hook + capture policy. The standing policy now nudges remember_preference after a correction, and the opt-in Stop hook scans the session for behavioural-correction phrases ("from now on…", "no, don't…", "always use…"): if you corrected the agent and no preference was saved, it blocks the stop once to ask — a second, independent reason alongside the existing "code changed but nothing logged" check.

v1.3.2

• Docs: added a "Using it day to day" section — the natural-language prompts that map to each tool, the one habit that matters ("have we hit this before?"), and the escape hatches. Clarifies that you talk to the agent rather than calling tools directly.

v1.3.1

• Stop hook: count direct memory edits as capture. The hook previously recognized only mcp__project-memory__* tool calls, so editing AGENTS.md / issues.jsonl directly (an endorsed capture path) still triggered the nag. It now also treats an Edit/Write to a file ending in AGENTS.md or issues.jsonl as captured — eliminating the false positive.
• append_decision/append_learning: no more duplicate sections. Heading matching was whole-line (^## Learnings$), so a heading with trailing text (## Learnings (gotchas …)) wasn't found and a duplicate section got appended. Now matches the heading's leading word.

v1.3.0

• Guaranteed capture (opt-in Stop hook). New install-hook / uninstall-hook subcommands register a Claude Code Stop hook that forces a single capture pass when a session changed code but recorded nothing to memory — turning the best-effort policy into a hard guarantee. Off by default, fires at most once per turn (no loops), silent when nothing changed or memory was already written, and disablable per-session via PROJECT_MEMORY_HOOK=off.

v1.2.0

• Sharper issue search. search_issues now matches only the text fields (symptom/cause/fix/id/tags) instead of the raw JSON, so queries no longer get false hits on field names. Added an optional tags filter; query is now optional (search by tags alone).
• sync_registry. Reconciles the root AGENTS.md projects table with the projects on disk — adds stub rows for projects missing from the table, flags rows whose directory is gone, and reports live open-issue counts. Automates the previously manual "new project → add a row" step. Hand-curated columns are preserved; apply=false reports drift only.
• find_by_file. Given a file path/fragment, returns the issues (via their files field) and the decisions/learnings (via AGENTS.md bullets that mention it) touching that file — code↔memory linking for "why is this code the way it is?".

v1.1.1

• Docs only: publishes the changelog to the npm page for parity (no functional change).

v1.1.0

• Proactive capture. The agent now records memory on its own instead of waiting for "log this": a standing capture policy is sent on initialize and the write/search tool descriptions are directive. It stays confirming (tells you what it logged), asks when unsure, and skips trivia/secrets. Explicit calls still work as an override.
• Docs: added "How it works (after install)" and "Proactive capture" sections.

v1.0.1

• Fix npx … install failing with "command not found" — the bin is renamed to project-memory-mcp to match the unscoped package name (npx resolution rule).

v1.0.0

• Initial release: stateless MCP server over AGENTS.md + issues.jsonl, 9 tools (project memory + issue tracking), npx … install for Claude Code and Cursor, and the push/pull memory model.