You (in Claude Code):  "Don't use any type in TypeScript"
  └─ amem stores this as a correction (priority 1.0, confidence 100%)

You (switch to Copilot): starts coding
  └─ Copilot already knows — amem feeds it the same correction

You (open Cursor): "What do you remember about TypeScript?"
  └─ Instantly recalls: "Don't use any type" + all related preferences

npm install -g @aman_asmuei/amem
amem-cli init      # Detects & configures all installed AI tools
amem-cli rules     # Generates extraction rules for proactive memory use

Or add to your MCP config manually:

{
  "mcpServers": {
    "amem": {
      "command": "npx",
      "args": ["-y", "@aman_asmuei/amem"]
    }
  }
}

The same engine powers amem (MCP server), aman-agent (CLI), aman-tg (Telegram bot), and any Node app you give memory to. Retrieval improvements ship via amem-core. MCP-tool changes ship via amem. They version independently.

The 97.8% R@5 headline is the engine quality from amem-core (LongMemEval-S, session-level, 500 questions, zero API calls) — exactly what you get whether you call it through MCP or import the library directly.

Memory Tiers

Temporal Validity

Memories aren't forever. When facts change:

• Old memories get expired (not deleted) — preserved for "what was true in March?"
• Contradictions are auto-detected ��� storing a new decision auto-expires the old one
• Query any point in time with memory_since

Your memory doesn't just store — it learns from its own structure. Call memory_reflect to trigger the reflection engine:

memory_reflect → Analyzes your entire memory graph
  │
  ├─ Clusters related memories (HNSW neighbor graph)
  ├─ Detects contradictions (negation pairs, numerical, low-overlap)
  ├─ Identifies synthesis candidates
  ├─ Surfaces knowledge gaps (topics with sparse recall)
  └─ Returns a structured report with suggested actions

The evolution loop:

1. Reflect — memory_reflect clusters your memories and finds patterns
2. Synthesize — AI merges related clusters into higher-order principles via memory_store
3. Link — memory_relate connects syntheses to source memories (tracked via synthesis lineage)
4. Repeat — each cycle, the graph becomes more coherent and abstract

The system auto-nudges when reflection is due (>7 days or >50 new memories since last run).

# Memory Reflection Report
Analyzed 127 memories in 12ms
Health Score: 68/100

## Stats
- Clusters: 8 (avg size: 4.2)
- Clustered: 34 | Orphans: 93
- Contradictions: 2
- Synthesis candidates: 3
- Knowledge gaps: 4

## Contradictions Found
⚠ Opposing language detected (23d apart, 87% similar)
  A: a1b2c3d4 "Always use semicolons in JavaScript..."
  B: e5f6g7h8 "Never use semicolons in JavaScript..."
  → Expire older memory a1b2c3d4 — newer supersedes it

## Synthesis Candidates
### cluster-0 (4 patterns)
  "These 4 related memories form a cluster about 'typescript, types':
  [patterns]:
    - 'Always use strict TypeScript types'
    - 'Prefer strict null checks'
    - 'Use unknown instead of any'
    - 'Enable strictNullChecks in tsconfig'

  Synthesize into a higher-order principle..."

## Knowledge Gaps
- "kubernetes deployment" — asked 3x, avg 25% confidence
- "database migration strategy" — asked 2x, avg 0% confidence

📈 Benchmarks

Recall Accuracy (LongMemEval)

All numbers from amem-core v0.5.1 — the retrieval engine powering this MCP server. Zero API calls, all local, fully reproducible.

Pipeline: local bge-small-en-v1.5 bi-encoder + ms-marco-MiniLM-L-6-v2 cross-encoder (int8, batched, default-on). See amem-core benchmarks for full per-type breakdowns, pipeline evolution, and honest notes.

Why this matters for the "rewrite it in Rust" question. The 10.3ms rerank figure above reflects a ~30% speedup over the per-pair implementation it replaced — achieved with ~20 lines of batching plus int8 quantization, no native rewrite. The hot paths were already efficient; the remaining wins came from using them more carefully. We stay on TypeScript.

Search Latency

🛠️ Tools Reference

Core Memory (7 tools)

📖 Usage Guide

Storing Memories

"Remember: we use PostgreSQL, not MongoDB"
"Store a correction: never use console.log in production"
"Note that the auth module is in src/auth/"

memory_store({
  content: "Never use 'any' — define proper interfaces",
  type: "correction",
  tags: ["typescript"],
  confidence: 1.0
})

Recalling Memories

// Step 1: Compact index — ~50-100 tokens (default)
memory_recall({ query: "auth decisions", limit: 5 })
// -> a1b2c3d4 [decision] Auth service uses JWT tokens... (92%)
// -> e5f6g7h8 [correction] Never store tokens in localStorage... (100%)

// Step 2: Full details only for what you need
memory_detail({ ids: ["a1b2c3d4", "e5f6g7h8"] })

// Multi-strategy: semantic + FTS5 + graph + temporal
memory_multi_recall({
  query: "authentication architecture",
  limit: 10,
  weights: { semantic: 0.4, fts: 0.3, graph: 0.15, temporal: 0.15 }
})

// Exact keyword search (FTS5 syntax)
memory_search({ query: "OAuth PKCE" })
memory_search({ query: '"event sourcing"' })     // phrase match
memory_search({ query: "auth* NOT legacy" })      // boolean

Managing Memories

// Surgical edit with auto-snapshot for rollback
memory_patch({ id: "a1b2c3d4", field: "content", value: "Updated text", reason: "clarified" })

// View edit history / restore
memory_versions({ memory_id: "a1b2c3d4" })

// Expire (preserve for history, exclude from recall)
memory_expire({ id: "a1b2c3d4", reason: "Migrated to GraphQL" })

// Promote to core tier (always loaded at session start)
memory_tier({ id: "a1b2c3d4", tier: "core" })

// Link related memories (graph builds itself, but you can add manual links)
memory_relate({ action: "relate", from_id: "abc", to_id: "xyz", relation_type: "supports" })

Reminders

reminder_set({ content: "Review PR #42", due_at: 1743033600000, scope: "global" })

reminder_check({})
// -> [OVERDUE] Review PR #42
// -> [TODAY] Deploy auth service
// -> [upcoming] Write quarterly report

reminder_complete({ id: "a1b2c3d4" })

Privacy

// Private blocks stripped before storage
memory_store({
  content: "DB password is <private>hunter2</private>, connect to prod at db.example.com",
  type: "topology", tags: ["database"]
})
// Stored: "DB password is [REDACTED], connect to prod at db.example.com"

// API keys, tokens, passwords auto-redacted by pattern matching
// Configure patterns in ~/.amem/config.json

⚔️ Honest Comparison: amem vs graphify

🌐 Platform Compatibility

Claude Code has the deepest integration (plugin + hooks + auto-memory sync). Copilot CLI is a close second. Other MCP clients get the full 33-tool server via manual config.

AI Skills

🔄 Working with Claude Code Auto-Memory

amem complements Claude's built-in auto-memory — it doesn't replace it.

Recommended: Keep both enabled. Run amem-cli sync to import Claude's memories into amem for unified, structured access.

amem-cli sync              # Import all projects
amem-cli sync --dry-run    # Preview what would be imported
amem-cli sync --project myapp  # Import specific project

amem-cli sync --to copilot              # Export to current project
amem-cli sync --to copilot --dry-run    # Preview without writing
amem-cli sync --to copilot --project /path/to/repo

Claude Code → amem sync → amem DB → amem sync --to copilot → copilot-instructions.md

📊 Dashboard & Knowledge Graph

amem-cli dashboard              # Opens at localhost:3333
amem-cli dashboard --port=8080  # Custom port

Full-featured web dashboard with:

• 🔍 Memory browser — search, filter by type/tier/source, inline actions (promote, demote, expire)
• 🕸️ Interactive knowledge graph — zoom, pan, click-to-focus with neighborhood highlighting, detail panel, search, directional edges
• 📈 Analytics — confidence distribution, type breakdown, session timeline
• ⏰ Reminders — view and manage cross-session tasks
• 📋 Copilot Preview — see what would be exported to copilot-instructions.md

💻 CLI Reference

# Setup
amem-cli init                          # Auto-configure AI tools
amem-cli rules                         # Generate extraction rules
amem-cli hooks                         # Install hooks for Claude Code
amem-cli hooks --target copilot        # Install hooks for GitHub Copilot CLI
amem-cli hooks --uninstall             # Remove hooks
amem-cli sync                          # Import Claude auto-memory → amem
amem-cli sync --to copilot             # Export amem → copilot-instructions.md
amem-cli doctor                        # Health diagnostics
amem-cli repair                        # Repair corrupted database from backups

# Dashboard
amem-cli dashboard                     # Web dashboard (localhost:3333)

# Memory operations
amem-cli recall "authentication"       # Semantic search
amem-cli stats                         # Statistics
amem-cli list --type correction        # List by type
amem-cli export --file memories.md     # Export to file
amem-cli forget abc12345               # Delete by short ID
amem-cli reset --confirm               # Wipe all data

🏗 Architecture

                        Your AI Tool
           Claude Code / Copilot CLI / any MCP client
                    │                │
                    │ MCP (stdio)    │ Lifecycle Hooks
                    ▼                ▼
          ┌─────────────────────────────────┐
          │   @aman_asmuei/amem             │  ← this package
          │                                 │
          │  33 Tools · 7 Resources · 2 Prompts
          │  Slash commands · CLI · Hooks   │
          │  Config: ~/.amem/config.json    │
          └────────────────┬────────────────┘
                           │ imports
                           ▼
          ┌─────────────────────────────────┐
          │   @aman_asmuei/amem-core        │  ← the engine
          │                                 │
          │  Multi-Strategy Retrieval       │
          │  [HNSW] + [FTS5] + [Graph] + [Temporal]
          │       + query expansion         │
          │       + cross-encoder reranker   │
          │                                 │
          │  Self-Evolving Reflection       │
          │  [Clustering] + [Contradictions]│
          │  + [Synthesis] + [Gap Detection]│
          │                                 │
          │  Embeddings: bge-small-en-v1.5  │
          │  Reranker: ms-marco-MiniLM int8 │
          │  97.8% R@5 on LongMemEval-S     │
          └────────────────┬────────────────┘
                           │
                           ▼
          ┌─────────────────────────────────┐
          │   SQLite + WAL + FTS5           │
          │   ~/.amem/memory.db             │
          │                                 │
          │   memories       (tiered)       │
          │   conversation_log (raw)        │
          │   memory_versions (history)     │
          │   memory_relations (graph)      │
          │   synthesis_lineage             │
          │   knowledge_gaps                │
          │   session_summaries             │
          │   reminders                     │
          └─────────────────────────────────┘

The amem MCP server is a thin wrapper around amem-core. The retrieval engine, embeddings, knowledge graph, reflection — all live in amem-core and version independently. Bug in MCP wiring? Republish amem. Recall improvement? Republish amem-core. No coupling.

Ranking Formula

score = relevance x 0.45 + recency x 0.2 + confidence x 0.2 + importance x 0.15

Additive scoring ensures no single low factor kills the ranking.

⚙️ Configuration

{
  "retrieval": {
    "semanticWeight": 0.4,
    "ftsWeight": 0.3,
    "graphWeight": 0.15,
    "temporalWeight": 0.15,
    "rerankerEnabled": true
  },
  "privacy": {
    "enablePrivateTags": true,
    "redactPatterns": ["..."]
  },
  "tiers": {
    "coreMaxTokens": 500,
    "workingMaxTokens": 2000
  },
  "hooks": {
    "enabled": true,
    "captureToolUse": true,
    "captureSessionEnd": true
  }
}

🧰 Tech Stack

🤝 Contributing

git clone https://github.com/amanasmuei/amem.git
cd amem && npm install
npm run build   # zero TS errors
npm test        # 281 tests pass

PRs must pass CI before merge. See Issues for open tasks.

_{Built with ❤️ in 🇲🇾 Malaysia by Aman Asmuei}

_{MIT License · Star ⭐ if amem saves your AI from amnesia}