Skip to content

examples

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

examples

README.md

Engram — Example Configs

Six self-contained examples, each with a README.md, bootstrap.py, and .env.example. Pick the one closest to your use case, run bootstrap.py, and you have a seeded Engram collection to build on.

Which config should I use?

# Config When to use
1 single-agent-personal-memory One agent, one user. The starting point for most personal AI assistants.
3 long-cycle-reflection-heavy Agent that runs for weeks/months and should get smarter over time (Reflection + Dream Engine).
4 multi-agent-shared-memory Two or more agents (e.g. Siri + BMO) sharing one Engram server with physically isolated collections.
5 claude-code-mcp-integration Wire Engram as the persistent memory backend for a Claude Code or any other MCP client.
6 qdrant-cloud-production Ship to production with Qdrant Cloud (TLS), HTTP REST + MCP simultaneously, and OTel tracing.
7 chatbot-session-memory Chatbot that needs short-lived per-user context (auto-expires in days, not months).

Numbers follow the internal Config numbering in /Engram/example-configs.md; Config 2 (episodic research agent) is in the backlog.

Config Examples

Unlike the bootstrap examples above (which seed memories), these three are runtime tuning profiles — each is a config.yaml + README.md you apply via the memory_apply_config MCP tool to shape how memories are ranked (retrieve_config) and written/deduped (update_config) for a given deployment shape.

Config When to use
config-personal-agent Single high-value, low-volume agent. Recall of a rare-but-critical fact matters more than precision (召回 > 精确).
config-team-knowledge Shared knowledge base, many contributors. Dedup + provenance matter more than exhaustive recall (精确 > 召回).
config-research-dedup Dense research notes with high semantic overlap. The most aggressive dedup profile — forces consolidation over fragment pile-up.

对比速查

维度 1. 个人 agent 2. 团队知识库 3. 研究笔记去重
主要风险 稀有重要记忆静默丢失 重复条目 + provenance 丢失 碎片堆积 + 反向优先级
取舍倾向 召回 > 精确 精确 > 召回 强制合并
dedupe_threshold 0.92（保守） 0.85（激进） 0.80（最激进）
recency_weight 0.20（低） 0.35（中高） 0.30（中）
importance_weight 0.35（高） 0.25（低） 0.20（最低）
min_score 0.55（宽） 0.65（紧） 0.70（最紧）
limit 6 10 8
特色开关 per-type 严 dedup provenance + require_source supersede 链 + 离线 consolidation

Some per-type / merge / consolidation sub-fields are marked # proposed in each config.yaml — intended productization knobs not yet all wired server-side. Each README has a ⚠️ proposed fields section flagging what is live vs roadmap.

Common prerequisites

All examples require:

  • Go 1.22+ (to build the engram binary) or Docker (for docker-compose up)
  • Qdrant — local (docker run qdrant/qdrant) or cloud (see Config 6)
  • One of:
    • OpenAI API key (ENGRAM_OPENAI_API_KEY=sk-...)
    • Voyage AI key (ENGRAM_VOYAGE_API_KEY=pa-...) — used in production

Each directory has its own .env.example; copy to .env and fill in your keys.

Quick recipe (any config)

# 1. Start Qdrant
docker run -d -p 6333:6333 -p 6334:6334 qdrant/qdrant

# 2. Build Engram
git clone https://github.com/FBISiri/engram.git
cd engram
go build -o engram ./cmd/engram/

# 3. Pick an example, configure, seed
cd examples/single-agent-personal-memory    # ← swap for your chosen config
cp .env.example .env && $EDITOR .env
pip install requests python-dotenv
python bootstrap.py

# 4. Run integration test to confirm everything works
cd ../..
ENGRAM_OPENAI_API_KEY=sk-... ./integration_test.sh

What bootstrap.py does

Each script seeds a curated set of example memories (typically 8–15) into the Qdrant collection via Engram's HTTP API (POST /memories). This means:

  • Memories get real server-side embeddings (not placeholder zeros)
  • Engram's TTL matrix sets automatic expiry based on type + importance
  • Engram's dedup gate (≥0.92) prevents duplicate seeding — re-running is safe

After bootstrapping, your collection is a live, search-ready starting point. Replace the example memories with your agent's real data as it accumulates.

Related docs