Hippocampus

---

Hippocampus is an MCP memory server you run yourself. It stores memories as plain Markdown files, indexes them in SQLite, and finds them with hybrid search: BM25 keywords + vector embeddings + a knowledge graph of links that you (or Claude) write between notes. Claude Code hooks recall relevant memories on every prompt and curate the memory at the end of each session.

Why Hippocampus?

Claude Code keeps memory per project folder. Open a session in another folder, or on another machine, and everything learned elsewhere is gone: which server runs what, how to reach the production database without being blocked, what was already fixed last week. You end up explaining the same things over and over, and Claude keeps retrying approaches that already failed.

Hippocampus moves those memories to one memory server you own, on your own computer or on a small VPS, and wires it into Claude Code, so that every session:

• recalls automatically. A hook searches memory on every prompt and injects the few notes that matter, before Claude starts working.
• writes to the shared memory. Claude gets MCP tools to search, read, save and update memories, and they are visible everywhere at once.
• tracks what is still open. Each memory has a status (active · pending · resolved · superseded), so "what's still pending?" has a real answer, and resolved items stop showing up as open.
• curates itself. At the end of a session an optional hook reviews the transcript, closes pending items that got done and saves new durable facts.

Memories are plain Markdown files. The database is only an index you can delete and rebuild at any time. Embeddings run locally (ONNX): no API key is needed and your notes never go to a third party.

Example: what Claude receives

Claude calls search_memory("restic backup on nas-1 is failing, disk full?") and gets this back (real output from the sample memories, trimmed):

# Relevant memories for: restic backup on nas-1 is failing, disk full?

## backup-restic-nas  (reference | project infra | 2026-01-20 | via meaning+name+text)
_Nightly restic backups to nas-1: schedule, retention, and how to restore a single file._
...
Disk usage history: [[nas-disk-full-march]].

## nas-disk-full-march  (project | project infra | 2026-03-20 | RESOLVED on 2026-03-21 | via graph neighbour)
_nas-1 hit 98% disk in March: old restic snapshots were never pruned. Fixed by adding prune to the timer._
...

> Old facts may be stale: check the date before acting on them.

The first note was found by keywords and meaning. The second came in through the [[nas-disk-full-march]] link written in the first one, and it is labelled as a resolved incident, so Claude treats it as history, not as the current state. The recall hook injects the same kind of block on every prompt, in a lighter form (fewer results, no link expansion) because it is paid on every message.

Other things you can say in any session, on any machine:

• "What's still pending on the shop API?" → list_pending
• "Save to memory how we fixed the orders report." → save_memory
• "The disk issue is fixed, mark it resolved." → mark_memory

How it works

flowchart LR
    subgraph Laptop["Your machines"]
        CC["Claude Code"]
        H1["recall hook<br/>(every prompt)"]
        H2["sync hook<br/>(end of turn)"]
        H3["extract hook<br/>(end of session)"]
    end
    subgraph Server["Your server"]
        API["Hippocampus<br/>MCP + REST + web"]
        MD[("Markdown files<br/>source of truth")]
        IDX[("SQLite index<br/>FTS5 + vectors + links")]
    end
    CC -- MCP tools --> API
    H1 -- /api/search --> API
    H2 -- /api/import --> API
    H3 -- "claude -p + MCP" --> API
    API --> MD --> IDX

A search mixes three signals:

1. BM25 (SQLite FTS5) for exact names, ports, flags and numbers;
2. vectors (paraphrase-multilingual-MiniLM-L12-v2, 384 dims, per section rather than per file) for meaning, in 50+ languages;
3. the link graph: each [[other-memory]] a human or Claude wrote adds a one-hop expansion that similarity search cannot infer.

The first two are fused with Reciprocal Rank Fusion. Resolved and superseded memories are demoted but still findable. The reasoning behind each choice is in docs/design.md.

How this differs from a vector-database memory

A memory that is only "embed every note, return the nearest vectors" misses several things that matter for an agent working on real systems:

	vectors only	Hippocampus
Exact tokens (5432, --force, a hostname)	often lost in the embedding	BM25 (FTS5) ranks them directly
Long notes	one vector per note blurs its topics	one vector per ## section
"This depends on that"	cannot be inferred from similarity	hand-written [[links]] expand results one hop
Outdated facts	returned as if still true	resolved / superseded demoted and labelled with a date
Open work	not modelled	pending status and list_pending
Storage	opaque database	Markdown files; the index can be rebuilt with hippocampus sync --full
Embeddings	often a hosted API	local ONNX model, no API key

Screenshots

Search results with the signals behind each hit	A memory with its links in and out

The web UI shows the whole memory as a constellation, which you can search, open, edit and filter by project.

Quick start

1. Choose where it runs

	On your computer	On a server (VPS)
Good for	one machine, trying it out	several machines, phone, claude.ai
You need	Python 3.10+	a small VPS (1 GB RAM + 2 GB swap) and a domain
Memory reachable from	this computer only	everywhere you use Claude
Setup	~5 minutes	~15 minutes

You can start local and move to a server later: memories are plain files, so you just copy the memories folder.

2a. On your computer

git clone https://github.com/marcelinollima/hippocampus && cd hippocampus
pip install .
hippocampus init        # creates ~/.hippocampus with a data folder and a token
hippocampus autostart   # starts the server now and at every login (no terminal to keep open)
python clients/claude-code/install.py --url http://127.0.0.1:8765

That's all: open a new Claude Code session. The web UI is at http://127.0.0.1:8765, and it asks for the token saved in ~/.hippocampus/server-token. To load the sample memories, copy examples/memories/* into ~/.hippocampus/data/memories/.

autostart uses the Startup folder on Windows, a LaunchAgent on macOS and a systemd user unit on Linux. None of them needs admin rights, and hippocampus autostart --remove undoes it. If you prefer Docker, use docker compose -f docker-compose.local.yml up -d (instructions are at the top of that file).

2b. On a server

git clone https://github.com/marcelinollima/hippocampus && cd hippocampus
cp .env.example .env         # set DOMAIN and HIPPOCAMPUS_TOKEN
mkdir -p data                # memories + index live here (back it up)
docker compose up -d         # Hippocampus + Caddy with automatic HTTPS

If you don't use Docker, see deploy/systemd. Then, on every machine where you use Claude Code:

python clients/claude-code/install.py --url https://memory.example.com --token <TOKEN>

The installer registers the MCP server (user scope) and the three hooks. It backs up ~/.claude/settings.json first, and --uninstall reverses everything. Your existing Claude Code memory folders (~/.claude/projects/*/memory) are uploaded on the next sync.

To use the same memory from claude.ai (web, desktop and mobile), see docs/claude-ai.md.

3. Use it

Nothing changes in how you work. Ask Claude something about your infrastructure and look for the <long-term-memory> block it received. To teach it something, say "save this to memory".

MCP memory tools

The server speaks MCP over streamable HTTP, so any MCP client that supports it can use these tools. The automatic recall, sync and curation hooks are specific to Claude Code.

tool	what it does
search_memory(query, limit)	hybrid search; returns a context block ready to use
read_memory(name)	the full memory, with its links in and out and similar memories
save_memory(name, description, body, type, project, status)	create or update; points to pending items that look alike
mark_memory(name, status, note)	change the status and append a dated note; same hint
list_pending(project, limit, offset)	what is still open, oldest first, paged
list_memories(project)	browse by project

Memory format

The format is the one Claude Code already uses for its own memory files, so existing notes work as they are:

---
name: shop-api-deploy
description: "How to deploy the shop API: build, copy, restart, check."
metadata:
  type: reference          # project | reference | feedback | user
  status: active           # active | pending | resolved | superseded
---

## Steps
1. `python -m compileall src` first: one SyntaxError takes down every route.
...
Related: [[shop-api-db-access]]

Folder = project: data/memories/<project>/<name>.md. You can edit files on disk and run hippocampus sync, or use the web UI, or let Claude do it.

Configuration

All server settings are environment variables. .env.example documents each of them. The most useful ones:

variable	purpose
HIPPOCAMPUS_TOKEN	bearer token for MCP, API and web UI (required)
HIPPOCAMPUS_OWNER	your name, used in the instructions Claude receives
HIPPOCAMPUS_LANG	en (default) or pt: language of the MCP instructions and of what the server writes into Claude's context
HIPPOCAMPUS_PINNED	a memory injected on top of every recall, e.g. a map of "nickname → repo, server, database"
HIPPOCAMPUS_URL_SECRET	enables /mcp/<secret> for clients that can't send headers, such as claude.ai custom connectors (guide)
HIPPOCAMPUS_MODEL	any fastembed text model

The client side lives in ~/.hippocampus/client.json, which is written by the installer. There you can map local project folders to server projects, add extra memory folders, or turn off automatic extraction. Install with --lang pt to get the recall header and the automatically extracted memories in Portuguese. The web UI follows your browser's language and has an EN/PT button to switch.

Security

Memories tend to contain exactly what an attacker wants: hostnames, database names and access recipes. Hippocampus is built around that:

• one bearer token, compared in constant time, on every route except / and /health;
• the app listens on 127.0.0.1 only, behind a TLS proxy (Caddy);
• DNS-rebinding protection on /mcp (explicit list of allowed hosts);
• recalled memories are labelled as notes, not instructions, and the extraction prompt treats the transcript as evidence only.

Read SECURITY.md before exposing a server, and never commit your data/ folder.

CLI

hippocampus init                  # prepare a local install (~/.hippocampus + token)
hippocampus autostart [--remove]  # run the server at every login (Windows, macOS, Linux)
hippocampus serve                 # MCP + API + web (syncs on start)
hippocampus sync [--full]         # index new/changed files, or rebuild everything
hippocampus suggest-links         # compute "similar" edges for the constellation
hippocampus stats                 # counts, broken links, pending items
hippocampus token                 # generate a random token

Pending items pile up when the fix is written somewhere else. To review them, run the Claude Code client's reviewer (it uses claude -p, with no tools):

python ~/.hippocampus/hooks/review.py            # report: which pending items look done
python ~/.hippocampus/hooks/review.py --apply ~/.hippocampus/review-<date>.json

It only marks what the report lists as resolved with a quote as proof, and you can veto any item by setting "apply": false in the .json first.

On a server, the review can run by itself once a day (deploy/systemd/hippocampus-review.service), with Claude Code under your subscription. It only looks at items that changed or got newer evidence since their last review, and HIPPOCAMPUS_REVIEW says what to do with the result: off (default), report (list_pending shows "looks done" with the quote, for a person to confirm) or apply (marks them resolved; undo by marking them pending again).

Project status

Hippocampus came out of its author's daily work, where it holds 360+ memories across 9 projects. It is young (0.x): expect rough edges and please report them. CI runs the tests on Linux and Windows, a Docker smoke test, a secret scan, and the full local install on Windows and macOS.

Known limitations:

• one token per server: it is a personal memory, not a multi-user one;
• automatic recall and curation hooks exist for Claude Code only; other MCP clients get the tools but must call them themselves;
• link expansion is one hop, over hand-written links only (by design, see docs/design.md).

Bug reports and ideas are welcome in the issues, and contributions in English or Portuguese. See CONTRIBUTING.md. Changes are listed in the CHANGELOG.

License

MIT