engram（engram 搜索）

Engram: Persistent Memory for AI Agents

Engram gives you durable semantic memory that survives sessions, compaction, crashes. All local, no cloud, no token cost.

Boot Sequence (MANDATORY)

On every session start, run:

engram search "<current task context>" --limit 10

Example: engram search "client onboarding status churn risk" --limit 10

This recalls relevant memories from previous sessions before you start work.

Storing Memories

5 memory types: fact | decision | preference | event | relationship

# Facts — objective information
engram add "API rate limit is 100 req/min" --type fact --tags api,limits

# Decisions — choices made
engram add "We chose PostgreSQL over MongoDB for better ACID" --type decision --tags database

# Preferences — user/client likes/dislikes
engram add "Dr. Steph prefers text over calls" --type preference --tags dr-steph,communication

# Events — milestones, dates
engram add "Launched v2.0 on January 15, 2026" --type event --tags launch,milestone

# Relationships — people, roles, connections  
engram add "Mia is client manager, reports to Danny" --type relationship --tags team,roles

When to store:

• Client status changes (churn risk, upsell opportunity, complaints)
• Important decisions made about projects/clients
• Facts learned during work (credentials, preferences, dates)
• Milestones completed (onboarding steps, launches)

Searching

Semantic search (finds meaning, not just keywords):

# Basic search
engram search "database choice" --limit 5

# Filter by type
engram search "user preferences" --type preference --limit 10

# Filter by agent (see only your memories + global)
engram search "project status" --agent theo --limit 10

Context-Aware Recall

Recall ranks by: semantic similarity × recency × salience × access frequency

engram recall "Setting up new client deployment" --limit 10

Better than search when you need the most relevant memories for a specific context.

Memory Relationships

7 relation types: related_to | supports | contradicts | caused_by | supersedes | part_of | references

# Manual relation
engram relate <memory-id-1> <memory-id-2> --type supports

# Auto-detect relations via semantic similarity
engram auto-relate <memory-id>

# List relations for a memory
engram relations <memory-id>

Relations boost recall scoring — well-connected memories rank higher.

Auto-Extract from Text

Ingest extracts memories from raw text (rules-based by default, optionally LLM):

# From stdin
echo "Mia confirmed client is happy. We decided to upsell SEO." | engram ingest

# From command
engram extract "Sarah joined as CTO last Tuesday. Prefers async communication."

Uses memory types, tags, confidence scoring automatically.

Management

# Stats (memory count, types, storage size)
engram stats

# Export backup
engram export -o backup.json

# Import backup
engram import backup.json

# View specific memory
engram get <memory-id>

# Soft delete (preserves for audit)
engram forget <memory-id> --reason "outdated"

# Apply decay manually (usually runs daily automatically)
engram decay

Memory Decay

Inspired by biological memory:

• Every memory has salience (0.0 → 1.0)
• Daily decay: salience *= 0.99 (configurable)
• Accessing a memory boosts salience
• Low-salience memories fade from search results
• Nothing deleted — archived memories can be recovered

Agent Scoping

4 scope levels: global → agent → user → session

By default:

• Agents see their own memories + global memories
• --agent <agentId> filters to specific agent
• Scope isolation prevents memory bleed between agents

REST API

Server runs at http://localhost:3400 (start with engram serve).

# Add memory
curl -X POST http://localhost:3400/api/memories \
  -H "Content-Type: application/json" \
  -d '{"content": "...", "type": "fact", "tags": ["x","y"]}'

# Search
curl "http://localhost:3400/api/memories/search?q=query&limit=5"

# Recall with context
curl -X POST http://localhost:3400/api/recall \
  -H "Content-Type: application/json" \
  -d '{"context": "...", "limit": 10}'

# Stats
curl http://localhost:3400/api/stats

Dashboard: http://localhost:3400/dashboard (visual search, browse, delete, export)

MCP Integration

Engram works as an MCP server. Add to your MCP client config:

{
  "mcpServers": {
    "engram": {
      "command": "engram-mcp"
    }
  }
}

MCP tools: engram_add, engram_search, engram_recall, engram_forget

Configuration

~/.engram/config.yaml:

storage:
  path: ~/.engram

embeddings:
  provider: ollama           # or "openai"
  model: nomic-embed-text
  ollama_url: http://localhost:11434

server:
  port: 3400
  host: localhost

decay:
  enabled: true
  rate: 0.99                 # 1% decay per day
  archive_threshold: 0.1

dedup:
  enabled: true
  threshold: 0.95            # cosine similarity for dedup

Best Practices

1. Boot with recall — Always engram search "<context>" --limit 10 at session start
2. Type everything — Use correct memory types for better recall ranking
3. Tag generously — Tags enable filtering and cross-referencing
4. Ingest conversations — Use engram ingest after important exchanges
5. Let decay work — Don't store trivial facts; let important memories naturally stay salient
6. Use relations — auto-relate after adding interconnected memories
7. Scope by agent — Keep agent memories separate for clean context

Troubleshooting

Server not running?

engram serve &
# or install as daemon: see ~/.engram/daemon/install.sh

Embeddings failing?

ollama pull nomic-embed-text
curl http://localhost:11434/api/tags  # verify Ollama running

Want to reset?

rm -rf ~/.engram/memories.db ~/.engram/vectors.lance
engram serve  # rebuilds from scratch

Created by: Danny Veiga (@dannyveigatx) Source: https://github.com/Dannydvm/engram-memory Docs: https://github.com/Dannydvm/engram-memory/blob/main/README.md