Agent Memory Hall

Your AI coding agent forgets everything the moment you close the session. AMH fixes that — with a handoff protocol, not another vector database.

License: Apache 2.0 MCP

---

The Problem

Every developer who uses AI coding agents knows this ritual:

1. Work with Claude Code / Cursor / Copilot for two hours
2. Close the session
3. Open a new session — the agent has no idea what you were doing
4. Spend 10 minutes re-explaining the context
5. Repeat

Worse: the agent said "I'll fix the auth bug next session." It didn't. Because it forgot it promised.

Existing memory tools solve storage ("where do I put memories?"). AMH solves continuity ("did the agent actually follow through?").

---

What AMH Does

Three slash commands. One local SQLite file. Three minutes to set up.

/start    → reads last session's state, shows "here's where you left off"
/save     → checkpoint mid-session (after a milestone)
/wrap-up  → structured close: what was done, what's pending, what's next

Plus Baton — the part no other tool does:

Baton tracks:
  - open_loops:      "said I'd do X — haven't done it yet"
  - follow_ups:      "deferred 3 times, first seen June 10"
  - active_decisions: "chose A over B because..."
  - patterns:        "keeps making the same mistake"

When you /start the next session, Baton tells your agent: "You promised to fix the auth bug. You haven't. It's been deferred twice."

---

Quick Start (3 minutes)

Option A: SQLite Skills (zero dependencies)

macOS:

mkdir -p ~/.claude/skills/{session-start,session-save,session-wrap-up} ~/.claude/job-memo
python3 -c "
import sqlite3, os
db = os.path.expanduser('~/.claude/sessions.db')
conn = sqlite3.connect(db)
conn.execute('''CREATE TABLE IF NOT EXISTS sessions (
    id INTEGER PRIMARY KEY AUTOINCREMENT,
    created_at TEXT DEFAULT (datetime(\"now\", \"localtime\")),
    slug TEXT NOT NULL,
    type TEXT NOT NULL DEFAULT \"episode\",
    content TEXT NOT NULL,
    tags TEXT DEFAULT \"\"
)''')
conn.commit(); conn.close()
print(f'Done: {db}')
"

Then copy the three SKILL.md files → macOS Guide | Windows Guide

Option B: MCP Server (one line)

npx @chibakuma/agent-memory-hall

Add to your MCP client (Claude Desktop / Cursor / Cline):

{
  "mcpServers": {
    "agent-memory-hall": {
      "command": "npx",
      "args": ["@chibakuma/agent-memory-hall"]
    }
  }
}

---

When You Outgrow SQLite → memory-hall Server

The SQLite entry point is enough for solo developers. When you need more:

Need	Solution
Semantic search ("find sessions about auth refactor")	memory-hall server — BM25 + vector hybrid
Multi-device sync	memory-hall server on your LAN or cloud
Team sharing	memory-hall's team-memhall namespace
CJK-native tokenization	memory-hall's jieba-based FTS5
Baton with CAS locking	memory-hall's Baton Store API

# Upgrade: one Docker command
git clone https://github.com/MakiDevelop/memory-hall
cd memory-hall && docker compose up -d
# Then point AMH to the server

The format is compatible — SQLite sessions import directly into memory-hall.

---

Architecture

┌─────────────────────────────────────────────────┐
│  agent-memory-hall (this repo)                  │
│  ┌───────────────────────────────────────────┐  │
│  │ Session Handoff Protocol                  │  │
│  │  /start → /save → /wrap-up               │  │
│  │  + Baton (open loops, follow-ups,         │  │
│  │    decisions, patterns)                   │  │
│  └───────────────────────────────────────────┘  │
│  ┌──────────────┐  ┌──────────────────────┐     │
│  │ MCP Server   │  │ SQLite Skills        │     │
│  │ (npx)        │  │ (SKILL.md, 0 deps)   │     │
│  └──────┬───────┘  └──────────────────────┘     │
│         │                                       │
│         ▼                                       │
│  ┌──────────────────────────────────────────┐   │
│  │ Storage: local SQLite (default)          │   │
│  │    or → memory-hall server (upgrade)     │   │
│  └──────────────────────────────────────────┘   │
└─────────────────────────────────────────────────┘
         │ optional
         ▼
┌─────────────────────────────────────────────────┐
│  memory-hall (separate repo)                    │
│  Hybrid search · Multi-device · Team sharing    │
│  CJK-native · Baton Store · HTTP/CLI/Python     │
│  github.com/MakiDevelop/memory-hall             │
└─────────────────────────────────────────────────┘

Two Repos, Clear Boundary

	agent-memory-hall (this repo)	memory-hall
What	Protocol + client + skills	Server engine
Solves	"How do sessions hand off?"	"Where do memories live at scale?"
Storage	Local SQLite (built-in)	SQLite + vector + hybrid search
Runs on	Your dev machine	Mac mini / Docker / cloud
Required?	Yes (entry point)	No (upgrade when needed)
MCP	Native (primary interface)	HTTP API (AMH connects to it)

---

Built-in Governance

AMH doesn't just store memories — it tracks who wrote it, where it came from, and whether it's trustworthy.

Feature	What it does
source_tier	raw_source / llm_derived / human_confirmed — trust level
trust_proof	Evidence-based verification for tier upgrades
provenance_chain	Full derivation history
content_hash dedup	BLAKE3 hash rejects duplicates
namespace isolation	Scoped read/write per project
audit trail	Immutable append-only log

Packages

Package	Description
@chibakuma/agent-memory-hall	Core protocol + MCP server + CLI
@chibakuma/aca-inspector	Web UI for memory governance
@chibakuma/aca-langgraph	LangGraph.js checkpointer integration
@chibakuma/aca-incident-analyzer	Governance linter (8 rules)
@chibakuma/aca-certification	Conformance test suite (5 layers)

What AMH Is Not

• Not a vector database. Use Pinecone / Chroma / Qdrant for that.
• Not a universal memory framework. Use Mem0 / Cognee / Letta if you need one.
• Not a replacement for CLAUDE.md. AMH complements file-based context injection.
• AMH solves continuity — ensuring nothing gets dropped between sessions.

---

Docs

• Tutorial: Why Your AI Assistant Needs Session Memory
• macOS Setup Guide
• Windows Setup Guide
• Full Protocol Spec

License

Apache 2.0 — by Cypher Lockhart (江中喬 / 千葉牧人)