Performance release. No public API changes — drop-in upgrade from 1.1.x.

Faster compression at scale

Compression (consolidation + contradiction arbitration) is now vectorised via an optional numpy accelerator. A full compression pass over an agent with 10,000 facts drops from ~118s to ~5s — about 23x.

pip install 'agentkeeper-ai[fast]' # enables the numpy accelerator

Without numpy, behaviour is unchanged: the pure-Python fallback is preserved, so the core keeps zero required dependencies.

Also in this release

• New [fast] extra (numpy), also bundled in [all].
• benchmark/stress_test.py — a reproducible scaling benchmark you can run yourself (10k-fact insert, 500 compression cycles, recall latency, save/load integrity, graph traversal).
• tests/test_fastmath.py — verifies the numpy and pure-Python paths produce identical results.
• Consolidation clustering now picks the best-matching centroid rather than the first above threshold (tighter clusters).

Built by Tom Anciaux Berner — ThinkLanceAI

Cognitive continuity infrastructure for long-lived AI agents.

AgentKeeper reconstructs an agent's full cognitive state — identity, memory, decisions, relationships — across model switches, crashes, restarts, and constrained context windows.

Install

pip install agentkeeper-ai

Highlights in 1.1

• Memory classes — decision(), preference(), constraint(), relationship(), task_state(), transient(), each with its own decay behaviour.
• Cognitive observability — agent.health() reports memory volume, importance distribution, contradiction count, stale ratio.
• GDPR-native retention — TTLs (ttl="30d"), gdpr_export() (Article 20), gdpr_purge() (Article 17).
• Persistent vector index — sqlite-vec backend survives restarts without re-embedding; scales to 10k+ facts.
• Pluggable storage — BaseStorage ABC, default SQLiteStorage, opt-in EncryptedSQLiteStorage (Fernet at-rest encryption).
• Async LLM consolidation — AsyncAgent.compress(use_llm=True) end-to-end.
• Graph memory — directed triples (agent.link), BFS traversal (agent.find_related), shortest path.
• Native MCP server — agentkeeper-mcp CLI for Claude Desktop, Claude Code, Cursor, and any MCP host. Nine tools exposed.
• Framework integrations — LangChain and CrewAI helpers, no hard dependency.

Quality

459 tests, ruff-clean, py.typed, CI on Python 3.10 / 3.11 / 3.12. Full backward compatibility with 1.0.

Built by Tom Anciaux Berner — ThinkLanceAI