memory_map

Persistent memory and conversation history MCP server for Claude Code — key-value context store, rolling MongoDB-backed history, and cross-project recall across sessions.

When to Activate

• Installing memory_map for the first time or on a new machine
• Writing CLAUDE.md session-setup instructions (load_memory, suggest_history)
• Deciding what belongs in memory vs history vs inline code comments
• Pruning stale or sensitive history chunks (delete_history)
• Using cross-project or global memory tools
• Configuring history compression or vector search
• Debugging why Claude starts a session without prior context
• Manually checkpointing conversation history with /mem_save

Architecture

memory_map_mcp/
├── server.py          — MCP server: 22 tools via FastMCP (stdio transport)
├── history_hook.py    — hook script: saves Q&A pairs on UserPromptSubmit / Stop / PreCompact
└── history_store.py   — storage layer: MongoDB CRUD, BM25, vector search, RRF, MMR

Storage:
├── MongoDB memory_map.history    — conversation chunks (suggest_history, save_history, delete_history)
├── MongoDB memory_map.memory     — key-value memory (save_memory / load_memory)
└── .mcp_memory.json              — per-project fallback when MongoDB not configured

MongoDB is required for all history features. Key-value memory works without it (falls back to .mcp_memory.json), but suggest_history, save_history, and delete_history all return an error if MEMORY_MAP_MONGO_URI is unset.

MCP registration makes all tools available globally. Per-project activation is controlled by CLAUDE.md — Claude only calls load_memory / suggest_history automatically if the instructions tell it to.

Installation

Step 1 — Install from PyPI

pip install memory-map-mcp

Or install from source for development:

git clone https://github.com/kid-sid/memory_map.git
cd memory_map
python -m venv venv

# Windows
venv\Scripts\pip install -e .

# Mac/Linux
source venv/bin/activate && pip install -e .

Step 2 — Set MongoDB URI

Add to your environment / shell profile:

export MEMORY_MAP_MONGO_URI="mongodb+srv://user:pass@cluster.mongodb.net/"

Free-tier MongoDB Atlas works. Without this, history tools are unavailable.

Step 3 — Register the MCP Server

# Global — available in every project (recommended)
claude mcp add -s user memory_map memory-map-mcp

If installed from source:

# Windows
claude mcp add -s user memory_map \
  C:/Users/yourname/memory_map/venv/Scripts/python.exe \
  C:/Users/yourname/memory_map/memory_map_mcp/server.py

# Mac/Linux
claude mcp add -s user memory_map \
  python3 /home/yourname/memory_map/memory_map_mcp/server.py

Registration scope options:

Scope flag	Stored in	Available
-s user	~/.claude/mcp.json	All projects on this machine
-s project	.claude/mcp.json	This repo only (committed, shared)
-s local	.claude/mcp.local.json	This repo only (gitignored, personal)

Always use -s user for memory_map — it stores files at local paths that differ per machine.

Verify: claude mcp list → should show memory_map.

Step 4 — Lifecycle Hooks

Add to ~/.claude/settings.json so history is captured automatically in every project:

{
  "hooks": {
    "UserPromptSubmit": [
      {
        "matcher": "",
        "hooks": [{
          "type": "command",
          "command": "memory-map-hook",
          "timeout": 10
        }]
      }
    ],
    "PreCompact": [
      {
        "matcher": "",
        "hooks": [{
          "type": "command",
          "command": "memory-map-hook --force",
          "timeout": 15
        }]
      }
    ],
    "Stop": [
      {
        "matcher": "",
        "hooks": [{
          "type": "command",
          "command": "memory-map-hook --force",
          "timeout": 15,
          "async": true
        }]
      }
    ]
  }
}

If installed from source, replace memory-map-hook with the full path to history_hook.py.

Hook	When	Flag
UserPromptSubmit	Every message — incremental saves	none
PreCompact	Before context window compaction	--force
Stop	When Claude finishes a turn	--force, async: true

Step 5 — Enable Per-Project Memory

Add this block to CLAUDE.md in the project root:

## Session Setup (Required)
At the start of every session, before doing anything else:
1. Call `load_memory` with the current working directory
2. Call `suggest_history` with the current working directory and the user's first message
3. Read both outputs before exploring files or asking questions

Save or update memory entries whenever you learn something worth keeping across sessions.
If something loaded from memory is no longer accurate, update it with `save_memory` using the same key.
Use short, lowercase keys: `stack`, `current_work`, `gotchas`, `key_files`. Keep values concise.

Commit CLAUDE.md — teammates get session-start loading automatically.

---

Memory Tools

save_memory / load_memory / delete_memory

Per-project key-value store backed by MongoDB (or .mcp_memory.json without MongoDB).

save_memory(project_path, key, value)
load_memory(project_path, query="", top_k=10)   → compressed key-value output
delete_memory(project_path, key)

Key conventions:

Key	What to store
stack	Language, framework, runtime versions
current_work	Active feature, bug, or initiative
gotchas	Non-obvious constraints, known failures, env quirks
key_files	Entry points, config files, critical paths
conventions	Non-obvious team decisions not in the code

Rules:

• Short, lowercase, underscore-separated keys
• Values: one or two sentences max — dense, not verbose
• Overwrite stale values with the same key; don't accumulate duplicates
• Convert relative dates to absolute: "Thursday" → "2026-05-15"

What NOT to save in memory:

• Code patterns and architecture (read the code)
• Git history / who changed what (git log / git blame)
• Fix recipes (the fix is in the code; commit message has context)
• Ephemeral task state (in-progress work, current conversation)
• Anything already in CLAUDE.md

Global Memory

Shared across all projects on the machine:

save_global_memory(key, value)
load_global_memory()

Use for: user identity, preferred tools, cross-project conventions. Never store secrets.

---

History Tools

Conversation history is stored in MongoDB (memory_map.history collection), one document per Q&A pair. Tags are extracted by local keyword matching — no LLM calls.

suggest_history — primary session-start tool

suggest_history(project_path, user_message, token_budget=2000, diversity=0.3)

Hybrid retrieval pipeline:

1. Concurrent fetch — vector search + BM25/tag scoring run in parallel
2. RRF merge — Reciprocal Rank Fusion combines both ranked lists
3. MMR re-ranking — penalises redundant chunks so results cover more distinct topics
4. Anchor — most recent chunk always included for session continuity
5. Token budget — fills remaining budget with most recent unselected chunks

Call at session start with the user's first message. Returns the most relevant chunks within the token budget.

save_history / load_history / get_history_chunks

save_history(project_path, dialogue, session_id="", tags="")
load_history(project_path, last_n=5)        → tag index (id, timestamp, tags, preview, tokens)
get_history_chunks(project_path, ids)       → full dialogue for comma-separated chunk IDs

save_history is called automatically by history_hook.py — no manual calls needed during normal use. Auto-summarises oldest chunks when total tokens exceed MCP_HISTORY_MAX_TOKENS (default 50 000).

load_history + get_history_chunks are for inspection only — listing what's saved or fetching a specific chunk by ID. Do not use them at session start — use suggest_history instead.

delete_history — prune stale or sensitive chunks

delete_history(project_path, ids="", older_than_days=0)

Remove chunks you no longer need. At least one filter must be provided:

# Delete specific chunks by ID (get IDs from load_history or suggest_history output)
delete_history("C:/projects/my-api", ids="6830a1f2e4b0c1234567abcd,6830a1f2e4b0c1234567ef01")
→ "deleted: 2 chunk(s)"

# Remove everything older than 30 days
delete_history("C:/projects/my-api", older_than_days=30)
→ "deleted: 7 chunk(s)"

# Both filters together (union — deletes chunks matching either condition)
delete_history("C:/projects/my-api", ids="6830a1f2e4b0c1234567abcd", older_than_days=30)

Deletion is scoped to the given project — other projects are never affected.

summarise_history

summarise_history(project_path, n=10)

Collapse the n oldest chunks into a single summary chunk. Triggered automatically by save_history when the token budget is exceeded; call manually to compact immediately.

backfill_history_embeddings / backfill_bm25_text

backfill_history_embeddings(project_path="", batch_size=20)
backfill_bm25_text(project_path="", batch_size=100)

Run once after enabling vector search or upgrading from an older version. Repeat until remaining=0. Pass project_path="" to backfill across all projects.

---

Vector Search (Optional)

Set MEMORY_MAP_EMBED_PROVIDER to enable semantic search in suggest_history:

Value	How
openai	text-embedding-3-small (1536-dim); requires OPENAI_API_KEY
local	sentence-transformers/all-MiniLM-L6-v2 (384-dim, CPU, no API key)
atlas	Atlas autoEmbed (Voyage-4) via Atlas Vector Search
(unset)	BM25 + tag scoring only — no vector search

Without MEMORY_MAP_EMBED_PROVIDER, suggest_history still works using BM25 and tag matching.

---

Cross-Project Tools

Tool	What it does
list_projects	List all projects that have saved memory under a base path
get_project_summary(project_path)	One-line summary of a project's memory
load_cross_project_memory(base_path)	Load memory from sibling projects
search_across_projects(base_path, keyword)	Full-text search across all project memories

Use cases:

• Referencing a pattern from a sibling project
• Finding which project owns a shared library
• Onboarding to a new project by comparing to a known one

---

Utility Tools

Tool	What it does
get_local_structure(path, max_depth=5)	Gitignore-aware directory tree
get_github_structure(repo, branch="main", max_depth=5)	GitHub repo file tree via API
get_git_history(path, count=5)	Recent commits as hash | subject

---

Compression

set_compression(project_path, level)

Level	Output	When to use
0	Raw — full fidelity	Debugging, inspecting memory content
1	Compact (default)	Normal use
2	Dense — abbreviations	Low context budget, large memories

Set per-project; persists until changed. Entries not updated in 30+ days get a [stale: Nd old] annotation at levels 1 and 2.

---

CLAUDE.md Session-Start Pattern

Minimal session-setup block for any project:

## Session Setup (Required)
At the start of every session, before doing anything else:
1. Call `load_memory` with the current working directory
2. Call `suggest_history` with the current working directory and the user's first message
3. Read both outputs before exploring files or asking questions

Save or update memory entries whenever you learn something worth keeping across sessions.
If something loaded from memory is no longer accurate, update it with `save_memory` using the same key.
Do not call `load_history` + `get_history_chunks` manually at session start — those are for inspection only.

---

Troubleshooting

Symptom	Likely cause	Fix
Claude starts sessions cold with no memory	CLAUDE.md missing or not instructing load_memory	Add session-setup block to CLAUDE.md
load_memory returns "no memory saved yet"	First session, or memory was deleted	Normal — save entries as context is established
History tools return "MongoDB unavailable"	MEMORY_MAP_MONGO_URI not set	Export the env var; verify with echo $MEMORY_MAP_MONGO_URI
suggest_history returns "no history yet"	Hooks not configured, or first session	Check hooks in settings.json; run memory-map-hook --force manually
MCP server not found	Not registered, or wrong path	Re-run claude mcp add; verify with claude mcp list
Hook hangs the session	No timeout on hook	Ensure "timeout": 10 is set on every hook
Stale chunks surfacing in suggestions	Old history never pruned	Run delete_history(project_path, older_than_days=N)

---

Red Flags

• Putting API keys or passwords in save_memory — memory may be committed or shared; store only key names and env var references, never values
• No CLAUDE.md after registering the MCP server — registration makes tools available but Claude only auto-loads memory if the session-setup instructions tell it to
• Using -s project for memory_map registration — memory_map stores files at local absolute paths that differ per machine; use -s user always
• Saving entire file contents or code blocks in memory — values should be one to two sentences; large values bloat the store and crowd out useful keys
• Hooks without timeout — a stalled hook blocks Claude Code indefinitely; always set "timeout": N
• Calling load_history + get_history_chunks at session start — use suggest_history instead; it runs hybrid retrieval and returns the most relevant chunks within a token budget
• Never running delete_history — history grows unbounded; prune stale projects periodically with older_than_days

---

Checklist

• pip install memory-map-mcp succeeded; memory-map-mcp command is on PATH
• MEMORY_MAP_MONGO_URI is set in the environment
• claude mcp list shows memory_map with correct entry point
• Registration used -s user (not -s project or -s local)
• Three lifecycle hooks in ~/.claude/settings.json: UserPromptSubmit, PreCompact, Stop
• Every hook has a "timeout" value; Stop hook has "async": true
• CLAUDE.md added to the project root and committed
• CLAUDE.md instructs load_memory and suggest_history at session start (not load_history)
• No secrets stored in save_memory — only descriptive values and env var names
• Memory keys are short, lowercase, and overwrite stale values (no duplicates)
• backfill_history_embeddings run if MEMORY_MAP_EMBED_PROVIDER was set after initial use
• /mem_save used before context compaction or ending a long session mid-task