handoff-mcp

Your agent's memory between sessions. Persistent, cross-project hand-off for Claude over the Model Context Protocol — it remembers the goal, the decisions, the dead-ends, and what's next, so the next session never starts from zero.

Claude forgets everything between sessions. handoff-mcp gives it a memory that survives the context window: it tracks the progress of your work — goals, decisions and their rationale, dead-ends, open questions, the next step — and at the boundary of a session produces a prioritised, token-budgeted brief so the next session resumes already knowing where you left off.

It is not a context dump. Two things make it different from similarity-based memory stores (mem0 / OpenMemory style):

1. Temporal supersession. A new decision retracts an old one; retracted decisions never appear in a future brief. Memory reflects the current state of the world, not a flat pile of contradictory facts.
2. Deterministic brief. The brief is assembled by an explainable ranker (recency + importance + supersession), not by an LLM — so it is reproducible and stays within a token budget (a soft cap on event content; see Limitations).
3. Your store, multi-device. The vault is your own private git repo, not a hosted service — so memory can follow you across machines (pull-on-start, push-on-checkpoint, conflict-free entity merges) with zero vendor lock. See Multi-device sync.

Status: v0.2, early. The core (vault, brief, supersession, keyword search, cross-project recall) is tested and stable; the semantic, consolidation, importer, and multi-device sync layers are optional and newer. It leans on the agent calling the tools at the right moments — see Limitations for the honest edges.

It is also cross-project: one vault, namespaced per project. The brief is project-scoped (where did I leave off here), but search_memory recalls across all projects — so when you say "in one of my projects we did X", Claude can find and pull that decision out of another project.

Contents

• Demo: two sessions
• Benchmarks
• How it works
• MCP tools
• Cost & API keys
• Semantic recall (optional)
• Memory consolidation (optional)
• Import existing history
• Quickstart
• Connect your MCP client — Claude, Cursor, Codex, Kilo Code, …
• Where your memory lives
• Multi-device sync (optional)
• Agent integration
• Limitations
• Development
• License

Demo: two sessions

python examples/two_sessions_demo.py — session 1 works and stops; session 2 is a fresh process that resumes from the brief alone.

Session 1 logs its progress (and changes its mind once):

log_event("goal",     "Ship the finance agent: income/expense tracking…")
d = log_event("decision", "Store transactions in a flat JSON file.")
log_event("decision", "Use SQLite instead of JSON — need queries.", supersedes=[d])  # retracts ↑
note_entity("Architecture", "SQLite-backed; LLM summary calls run in a worker.")     # durable
log_event("deadend",  "Provider streaming API times out on long months; needs chunking.")
log_event("question", "Recurring transactions: templates or materialised rows?")
log_event("next_step","Write the SQLite schema, then the ingest function.")
checkpoint("Chose SQLite; finance schema is next.")

Session 2 calls get_brief() and gets back only the current state — the retracted JSON decision is gone:

# Hand-off brief — agent-hub

## Goal
- Ship the finance agent: income/expense tracking with scheduled summaries.
## Next step
- Write the SQLite schema for transactions and categories, then the ingest function.
## Decisions
- Use SQLite for transactions instead of JSON — need queries for summaries. See [[Architecture]].
## Dead ends (tried & failed)
- Tried the provider's streaming API for the summary job — times out on long months. Don't retry without chunking.
## Open questions
- Should recurring transactions be modelled as templates or materialised rows?
## Related knowledge
- [[Architecture]] — SQLite-backed; LLM summary calls run in a worker.

The retracted "flat JSON file" decision never appears. The graph-linked [[Architecture]] note is pulled in automatically. The brief is ~170 tokens.

And cross-project recall — search_memory("timeout chunking", scope="all") pulls a decision out of a different project:

- [hermes]    In the Hermes project we solved long-job timeouts by chunking requests.
- [agent-hub] Tried the provider's streaming API for the summary job — times out on long months…

Benchmarks

Two offline, deterministic benchmarks — no LLM, no network — so they regenerate identically anywhere and are pinned by tests/test_benchmark.py. Full numbers, methodology, and how to reproduce: benchmarks/RESULTS.md.

1. Supersession in isolation (benchmarks/supersession_benchmark.py). A project's decisions evolve across 8 statements over 4 topics; each has one decision a later one retracts (JSON→SQLite, cookies→JWT, …). Retrieved by decision (so every stale fact is reachable), the flat log vs the active view:

Supersession removes exactly the retracted decisions while keeping every current one — and scoring current kept too means an empty answer can't pass as a win. A similarity store with no notion of one fact retiring another behaves like the OFF row.

2. Brief vs naive dump (benchmarks/brief_reconstruction.py). What a resuming session actually reads — the budgeted, supersession-aware brief vs pasting back the whole log. As history grows the dump balloons and keeps carrying every retraction; the brief stays bounded (a soft cap) and contradiction-free while retaining all key items (e.g. at 228 events: 3063→217 tokens, 14×, 3 contradictions → 0).

These measure the mechanism honestly rather than staging a head-to-head against another store — a fair cross-system run needs both under identical retrieval plus an LLM endpoint we can't reproduce in CI (see ADR-0008).

"Tokens" here and elsewhere in this README are estimated as len(text) / 4 (model-agnostic), not counted with a real tokenizer.

How it works

The markdown vault is the source of truth — human-readable, openable in Obsidian, your data on your disk. The SQLite + FTS5 index is derived from the vault and can be rebuilt at any time; it powers ranking, the token budget, and cross-project full-text search.

See docs/architecture.md and the ADRs for the design rationale.

MCP tools

Also exposed: an MCP resource session://brief and a prompt resume for auto-loading the brief at the top of a session.

log_event types: goal, decision, deadend, file, question, next_step.

Cost & API keys

No subscription. No required API keys. The core is free and fully local — your memory is plain files on your disk, and handoff-mcp never phones home (no hosted service, no telemetry).

So most of the value costs nothing and needs no key. The optional layers either run locally or use your own provider — you're never locked into ours.

Semantic recall (optional)

Keyword search (FTS5/bm25) is the default and needs nothing extra. You can enable a semantic layer that fuses keyword and embedding similarity with Reciprocal Rank Fusion:

HANDOFF_SEMANTIC=1 handoff-mcp     # turn the layer on (default: hashing backend)

The backend you pick decides whether this actually understands paraphrases. The default hashing backend is a lexical baseline — it hashes tokens, so it adds fuzzy lexical matching (and demonstrates the hybrid pipeline) but does not recall on meaning when the words differ. For genuine paraphrase-tolerant recall, choose local or openai, which use learned embeddings.

Pluggable embedding backends, selected with HANDOFF_EMBEDDER — one server, no forks:

# Example: semantic recall via any OpenAI-compatible endpoint
pip install -e ".[semantic-openai,semantic]"
export HANDOFF_SEMANTIC=1 HANDOFF_EMBEDDER=openai
export HANDOFF_EMBED_BASE_URL=https://your-openai-compatible-endpoint/v1 HANDOFF_EMBED_API_KEY=sk-…
export HANDOFF_EMBED_MODEL=text-embedding-3-small

Design (see ADR-0004):

• Pluggable embedder behind an Embedder protocol — the hashing default is deterministic and dependency-free; openai / local plug in for real semantic quality without changing anything else.
• sqlite-vec is an accelerator, not a requirement (.[semantic]) — vectors persist as BLOBs and search works with an exact cosine scan; if sqlite-vec is installed and loadable, a vec0 table provides fast KNN with the same top-ranked results.
• The deterministic brief never consults embeddings — semantics only affect search_memory.
• Reranking — recall results (any mode) are reordered by a deterministic blend of relevance + recency (time-decay) + importance, so fresh, high-priority memories surface first. No LLM; pass rerank=False to get raw relevance order.
• Incremental embedding — events are immutable, so startup only embeds new events (cached vectors are reused); the cache self-invalidates if the embedding model's dimension changes. This keeps heavier local models practical.

To use a lighter/faster local model instead, set HANDOFF_EMBED_MODEL (e.g. Alibaba-NLP/gte-multilingual-base, 305M/768-dim — also set HANDOFF_EMBED_TRUST_REMOTE_CODE=1 as that model requires it).

Memory consolidation (optional)

Over a long-lived project the episodic log grows without bound — a volume problem, not just an indexing one. Consolidation ("sleep") folds it down:

HANDOFF_LLM_MODEL=gpt-4o-mini handoff-mcp   # enables the consolidate tool

consolidate(project?, older_than_days?) distils old finished sessions' active decisions into the durable entity notes (Architecture, Decisions, Dead-ends, …), then archives the originals to <project>/archive/ and drops them from the active index. So the vault shrinks but the lasting knowledge — which the brief already surfaces — is kept.

• It is the only step that calls an LLM, and it's off unless HANDOFF_LLM_MODEL is set (OpenAI-compatible; reuses the embedder's endpoint settings). The brief, search, and supersession stay deterministic.
• Only active events are distilled — a retracted decision is never immortalised. Dead-ends are kept as cautionary facts.
• Originals are archived, not deleted (auditable, reversible).

See ADR-0006.

Import existing history

Bootstrap a project's memory from data you already have, so it's useful from minute one instead of empty:

handoff-import git ./my-repo --project my-project        # commit history → memory
handoff-import claude session.jsonl --project my-project # a Claude Code transcript

• git turns each commit into a decision (the subject) plus a files-touched note, timestamped at the commit date — fully deterministic, no LLM.
• claude pulls the first prompt as a goal and file edits as file events from a Claude Code transcript (best-effort).
• Import is idempotent (ids derive from the source), so re-running only adds what's new. Writes into the same shared vault (HANDOFF_VAULT).

Quickstart

uv venv && uv pip install -e ".[dev]"

# Run the two-session demo: session 1 works, session 2 reads the brief.
python examples/two_sessions_demo.py

Connect your MCP client

handoff-mcp speaks standard MCP over stdio, so it works with any MCP client — Claude, Cursor, Codex, Kilo Code, Windsurf, Cline, VS Code, Zed, … The config is essentially the same everywhere; only the file location (and, for Codex, the format) differs.

Canonical config — the mcpServers JSON block used by Claude, Cursor, Kilo Code, Windsurf, Cline, and most others:

{
  "mcpServers": {
    "handoff": {
      "command": "handoff-mcp",
      "env": {
        "HANDOFF_VAULT": "/path/to/your/vault",
        "HANDOFF_PROJECT": "my-project"
      }
    }
  }
}

Codex CLI uses TOML in ~/.codex/config.toml (or run codex mcp add):

[mcp_servers.handoff]
command = "handoff-mcp"
[mcp_servers.handoff.env]
HANDOFF_VAULT = "/path/to/your/vault"
HANDOFF_PROJECT = "my-project"

VS Code uses "servers" and an explicit type:

{ "servers": { "handoff": { "type": "stdio", "command": "handoff-mcp",
  "env": { "HANDOFF_VAULT": "/path/to/your/vault", "HANDOFF_PROJECT": "my-project" } } } }

Notes:

• handoff-mcp must be on PATH — install it as a tool (uv tool install git+https://github.com/kirill-sviridov/handoff-mcp; PyPI release coming) or, from a checkout, use "command": "python", "args": ["-m", "handoff_mcp.server"].
• Set HANDOFF_PROJECT per agent/repo; keep HANDOFF_VAULT pointed at the same shared vault across all of them (see below).

Where your memory lives

One central vault, with a folder per project inside it:

~/.handoff-mcp/vault/            # HANDOFF_VAULT (default; override per machine)
├── .index.db                    # derived SQLite index — rebuildable, gitignore it
├── my-project/
│   ├── sessions/<id>.md         # episodic notes
│   └── entities/<Name>.md       # durable knowledge
└── another-project/…

• One vault, not one-per-repo. Cross-project recall (search_memory) only works because every project lives in a single store. So point every agent's HANDOFF_VAULT at the same directory and just vary HANDOFF_PROJECT.
• It lives outside your code repos, so it never gets committed into your work projects by accident — your project repos stay clean.
• Want backup / multi-machine sync? The vault is plain markdown, so version it on its own (a private git repo, Obsidian Sync, Dropbox…). For built-in git sync across devices — pull-on-start, push-on-checkpoint, conflict-free entity merges — see Multi-device sync; handoff-sync --setup configures the remote and gitignores the derived index for you.
• Prefer memory that travels with one repo? Point HANDOFF_VAULT inside that repo (e.g. ./.handoff) — but then search sees only that project. The shared vault is recommended.

Multi-device sync (optional)

Your vault is just a private git repo, so memory can follow you across machines. Sync is strictly opt-in — with no git configured, memory works fully on one device.

Entity notes merge cleanly across machines via git's built-in union driver (*/entities/*.md merge=union in the vault's .gitattributes, written by setup). In a shell-less client (e.g. Cursor), just ask the agent to sync — the sync tool needs no terminal. Setup needs working git auth (gh auth login or an SSH key); if a push fails, the command tells you exactly what to fix.

Agent integration

Knowing when to use it. The server never pushes anything to the model; the model decides when to call the tools. Three layers make that reliable, from most portable to most capable:

1. Tool descriptions (built in) — every tool says when to call it. Works in any MCP client.

2. Server instructions (built in) — a short ritual the server sends on connect (load the brief at start, log as you work, checkpoint at the end). Portable across Claude Desktop, Cursor, and other MCP clients.

3. Claude Code skills (skills/) — encode the workflow and, crucially, trigger on colloquial cues:

   • session-handoff — when you say "го в следующую сессию" / "that's it for today", the agent knows to checkpoint on its own. It also bootstraps the project's instruction file on first use (CLAUDE.md, or AGENTS.md / .cursor/rules/handoff.mdc / .windsurfrules per client).
   • session-planning — optional companion: breaks a big task into session-sized chunks and persists the plan in memory.

   Install by copying into your skills dir:

   bash

   cp -r skills/handoff skills/planning ~/.claude/skills/   # user-wide
   # or .claude/skills/ inside a specific project
   

   (Skills are a Claude Code / claude.ai feature; other agents rely on layers 1–2.)

   For non-Claude clients, drop the memory block into the project's instruction file with the handoff-init CLI (idempotent):

   bash

   handoff-init                 # CLAUDE.md
   handoff-init --client codex  # AGENTS.md   ·   --client cursor / windsurf
   

For Claude Code specifically, also add this to your CLAUDE.md so the brief loads automatically even without the skill:

At the start of a session call get_brief. Record decisions, dead-ends and the next step with log_event as you work — one atomic item per call (1-2 sentences with the why), not a whole-session summary; reference durable notes as [[Entity]]. checkpoint before you stop. When I mention past work or another project, call search_memory.

Limitations

Honest edges of v0.2, so you know what you're adopting:

• Supersession is explicit, not inferred. handoff-mcp never decides on its own that one memory retires another — the agent must say so, via supersedes (by id) or supersedes_query (by best match). That is deliberate (it's what keeps the brief deterministic and auditable), but it means the quality of the memory depends on the agent actually logging retractions. It won't silently de-duplicate contradictions the way an LLM-extraction store attempts to.
• It depends on the agent's discipline. The server never pushes anything; value comes from the model calling log_event / get_brief / checkpoint at the right moments. The tool descriptions and the skill nudge this, but a client that never calls the tools gets an empty vault. Cross-session memory is only as good as what got logged.
• The default semantic backend (hashing) is lexical, not paraphrase-aware. Real paraphrase recall needs local or openai (see Semantic recall). The deterministic brief itself never uses embeddings.
• The token budget is soft. It bounds event content; section headings and the "Related knowledge" block are chrome on top, so the rendered brief can sit a little above the number. It keeps the brief bounded and flat as history grows — it is not a hard byte cap.
• Single-process freshness. One vault can be shared across processes/agents (SQLite WAL + a busy timeout let writers coexist), but a running process refreshes its view of the vault at startup (_sync_index) — it picks up its own writes live, but another process's new events only on the next launch (a sync — manual or the HANDOFF_AUTO_SYNC pull-on-brief — re-indexes mid-session when it pulls new events). For concurrent threads inside one process, access to the shared index is serialised by a lock.
• Personal/team scale. Ranking loads a project's events into memory; this is fine for thousands of sessions, not tuned for millions (see ADR-0005 for the localized fixes if that day comes).

Development

uv pip install -e ".[dev]"
ruff check . && mypy && pytest
python examples/two_sessions_demo.py        # the hand-off in action
python examples/demo_presentation.py         # paced/narrated version (for recording a GIF)
python examples/stdio_smoke.py              # run it as a real stdio MCP server
python benchmarks/brief_reconstruction.py   # brief vs naive full-dump
python benchmarks/supersession_benchmark.py # supersession on vs off, in isolation

License

MIT — see LICENSE.