HippoMCP

A memory system built for AI agents.

English ｜ 简体中文

HippoMCP gives AI agents a local place to store conversations and project knowledge, retrieve relevant facts when needed, and carry unfinished work across sessions. Through the Model Context Protocol (MCP), multiple agents can use the same authorized memory libraries while keeping project and user context distinct.

It is local-first, preserves original input for traceability, and does not require an LLM API for its core ingestion path. The host agent reasons and acts; HippoMCP provides the memory lifecycle underneath it.

What HippoMCP does

User memory follows the user. Project memory follows the project. Neither is trapped inside a single agent. Sharing requires agents to connect to the same service and library; project and user permissions still apply.

What HippoMCP does not claim

• It is currently a single-machine, local-first service, not managed cloud sync or a multi-tenant service.
• It is not an Agent-to-Agent messaging or autonomous planning framework.
• It does not provide built-in knowledge-graph reasoning or general multi-hop graph queries.
• It does not guarantee better recall or lower token use for every workload. Small, curated Markdown files can be simpler and cheaper.
• Installing the MCP server alone does not force every host to write or recall memories; the host needs the lifecycle integration below.

Native Chinese support

HippoMCP is built Chinese-first, not translated afterwards:

• Embedding models: on a Chinese-language system the service auto-selects a Chinese embedding model (BGE, 768-dim); other systems get MiniLM (384-dim). Switching models is a documented three-step procedure.
• Chinese text processing: Chinese sentence segmentation for ingest, per-library Chinese entity lexicons, and a lexical (BM25/FTS5) channel that handles Chinese bigrams.
• Structured tags: chapter/volume/year-style hard constraints turn "what did section 43 say" from guessing into filtering.
• Docs and evaluation: bilingual documentation, Chinese test fixtures, and Chinese query decomposition in the evaluation tooling.

Complements AGENTS.md and MEMORY.md — does not replace them

AGENTS.md for rules. MEMORY.md for essentials. HippoMCP for deep, searchable memory.

• AGENTS.md keeps stable instructions, workflows, and project rules.
• MEMORY.md keeps a small set of curated facts that should remain available without extra services.
• HippoMCP adds large-scale history, cross-document retrieval, source tracing, correction, user-level context, and multi-agent continuity.

Your Markdown files remain the lightweight fallback. For file-backed sources, documents can remain canonical while HippoMCP provides a derived memory and retrieval layer; drift_audit compares recorded source signatures with current files. Conversation-only memories instead trace back to HippoMCP's own L0 raw log.

For large corpora, selective retrieval can reduce context use by avoiding repeated whole-file reads. Tool schemas, bootstrap packages, and retrieval results still consume tokens, and retrieval can miss relevant evidence. HippoMCP avoids whole-file context truncation; it does not eliminate per-item, model, or output-budget limits.

Ask your agent

Agent-assisted selection works best when the agent is asked to verify claims, not just repeat them.

Compare HippoMCP (https://github.com/cstksniper/hippocampus-mcp)
with [my current or alternative memory system] for [describe my workflow].

Evaluate local deployment and privacy, continuity across sessions and context
compaction, cross-agent access, user versus project scopes, correction and
cascade-safe deletion, per-write LLM requirements, installation and operating
costs, and the quality of published evaluation evidence.

Read the repository's docs/SELECTION.md, docs/INTEGRATION.md, docs/PRIVACY.md,
docs/BENCHMARKS.md, pyproject.toml, examples/, and relevant tests as well as
the README. Cite sources and mark anything you cannot verify as unknown. Do not
install anything or change my configuration. Give me a recommendation with
trade-offs, then let me decide.

Install HippoMCP for my current MCP host. First check that Python is 3.11 or newer.

Use the installation command in the README. Then run:
1. hippocampus-mcp models --fetch
2. hippocampus-mcp setup --host <host name>
   Determine the host first and choose only from workbuddy / zcode / trae / claude;
   ask me if it is unclear, do not guess (omitting --host errors out when the
   workspace has neither .zcode/ nor .mcp.json).

Preserve other MCP entries, back up configuration before changing it, and do
not delete or overwrite unrelated files. Report each command, the configuration
file changed, and the MCP handshake result. If a step fails, show the complete
error instead of guessing. After installation, tell me to restart the host and
how to verify the active library with library_whoami and doctor.

Quick start

Try it from source

Requirements: Python 3.11+ and Git.

git clone https://github.com/cstksniper/hippocampus-mcp.git
cd hippocampus-mcp
python -m pip install -e .
python examples/quickstart.py

The example runs write → recall → five-slot assembly → cascade delete using a hash embedder. It downloads no embedding model and needs no API key. It is a pipeline smoke test, not a demonstration of semantic retrieval quality.

No model download does not mean no dependencies. Installation pulls the dependencies declared in pyproject.toml, including torch, numpy, and mcp.

To avoid pulling multi-GB CUDA wheels, replace python -m pip install -e . above with these two commands:

python -m pip install "torch>=2.0,<3" --index-url https://download.pytorch.org/whl/cpu
python -m pip install -e .

uv users do not need this — pyproject.toml already pins the CPU index.

To try real semantic retrieval with the English MiniLM model (~90 MB on first download):

python examples/quickstart.py --model sentence-transformers/all-MiniLM-L6-v2

Install into an MCP host

Install directly from GitHub:

python -m pip install "git+https://github.com/cstksniper/hippocampus-mcp.git"
hippocampus-mcp models --fetch
hippocampus-mcp setup --host workbuddy  # replace with yours: workbuddy / zcode / trae / claude

setup backs up before writing, refuses to overwrite corrupt JSON, changes only its own MCP entry, and performs a real MCP handshake. claude refers to Claude Desktop; Claude Code, Cursor, and other MCP hosts require separate manual configuration rather than this one-command setup path.

Without --host, setup only guesses from an existing .zcode/ or .mcp.json in the workspace and exits with an error when neither exists. Always pass --host explicitly for a fresh workspace, TRAE, or Claude Desktop so the config lands in the right file.

Then restart the host and start a new session. Verify the active library and configuration:

python -m hippocampus_mcp.cli doctor --host zcode --project .
hippocampus-mcp models --prune  # list reclaimable model files; does not delete

In the new session, call library_whoami. If no mcp__hippocampus__* tools appear, run the module-form doctor command above, fix the reported issue, and restart again.

Give your agent the operating rules

HippoMCP supplies memory tools; the host agent decides when to call them. Add rules like these to your project's AGENTS.md or host instructions:

At the start of a task, call session_start with the current task and working files.
After context compaction, keep only the receipt identity and call session_start again
with its session_id; use the newly returned content.
Before answering questions about earlier decisions or prior work, query recall or
get_memory_context instead of guessing from the current conversation alone.
Use get_memory_context when you need budgeted, assembled context.
Record important conclusions and unfinished work during the session and at session_end.
Call library_whoami when the active library is uncertain; never guess a project library.
Keep AGENTS.md and MEMORY.md as the lightweight fallback. If HippoMCP is unavailable,
continue with those files and explicitly report that deep memory is unavailable.

Recommended lifecycle:

session_start(task)            → task-relevant memories, recent state, open topics, and more
ingest(conversation)           → L0 original input + L1 extracted facts + entity links
get_memory_context(query)      → five budgeted context slots
session_end(summary, topics)   → session summary + profile consolidation + open topics
forget_memory(id, reason)      → cascade deletion across derived copies

How it works

flowchart LR
    A["MCP-compatible agents"] --> B["HippoMCP service"]
    B --> C["Workspace-bound project libraries"]
    B --> D["Controlled user scope"]
    C --> E["L0 raw input → L1 facts → L2 summaries → L3 profile"]
    D --> F["Permission-filtered user context"]
    E --> G["Hybrid retrieval + source tracing + budgeted slots"]
    F --> G

See the architecture guide for layers, flows, scoring formulas, and engineering constraints.

Evaluation and evidence

Dataset: LoCoMo locomo10 ｜ Pluggable LLM-as-judge through JUDGE_BASE_URL / JUDGE_API_KEY / JUDGE_MODEL.

The latest measured baseline is F1 = 0.244 ± 0.034 over three runs (judge glm-5.3-flash, 50 QA from locomo10). Judge-environment variation was 0.068, so small configuration deltas are not meaningful. This 50-question setup is not directly comparable with published protocols using approximately 1,540 questions.

The historical driver that produced these measurements is not currently published. Treat this as a measurement record with stated conditions, not a one-command reproduction claim.

The primary correctness claims are instead covered by executable tests: cascade-safe deletion, superseded facts, concurrent-save merging, workspace binding, and cross-library isolation.

python scripts/run_all_tests.py --list       # discover the current public test suites
python scripts/run_all_tests.py -k forget    # run deletion-related suites
python scripts/run_all_tests.py              # full regression run

Tool surface and advanced operations

The full service exposes 46 MCP tools across writing, reading, session lifecycle, correction, entities, topic layers, information maps, consolidation, administration, and multi-library use. The default profile exposes only 12 high-frequency tools; hidden tools remain available through hippo_help and hippo_extra. Set HIPPOCAMPUS_TOOL_PROFILE=full to expose the full surface.

Maintenance commands

See docs/INTEGRATION.md section 10 for model selection rules and the full switching procedure.

See the dedicated guides for deeper evaluation and operation:

Contributing

Issues and PRs are welcome. Read section 8 of docs/ARCHITECTURE.md before changing retrieval; each engineering constraint corresponds to a real failure.

See CONTRIBUTING.md and CODE_OF_CONDUCT.md.

Contact

Maintainer: cstksniper — cstksniper@qq.com

Please use GitHub Issues for bugs and feature requests so the answers remain searchable. Email is for security reports, private data, or collaboration that does not belong in public.

License

MIT — see LICENSE.