Hakuya — Provable Memory Infrastructure for AI Agents
Hakuya is open-source agent memory you can prove. Every other memory layer asks you to trust a black box: Hakuya records where each belief came from, lets it decay when it goes stale, can erase any subject on request, and writes every change to a tamper-evident audit trail — so you can prove what your agent knows, and why.
Trust & Governance
- Write-Provenance on Every Belief — Each memory carries its source, evidence type, and confidence, so you can always answer why the agent holds a belief.
- Tamper-Evident Audit Trail — Every change (created, reinforced, contradicted, redacted, erased) is appended to a per-tenant SHA-256 hash-chained log. Verify the whole chain in one call; export a signed record for SOC 2 / HIPAA review.
- Verified Per-Subject Erasure — Cryptographically shred everything about one customer, patient, or guest while keeping a provable record of the erasure. Built for GDPR Article 17 and the EU AI Act.
- Self-Cleaning Memory — Stale beliefs lose confidence; redundant ones suppress each other through competition-aware decay. Other stores accumulate junk; Hakuya's memory stays clean on its own.
The Cognitive Engine Underneath
- Four Memory Types — Semantic (facts, preferences), episodic (experiences with outcomes), procedural (learned skills), and working memory (active context).
- Belief Dynamics — Log-odds confidence updates: reinforced memories strengthen, contradicted ones weaken, unused ones decay. Confidence-based tiering auto-injects hot memories into agent context.
- Contradiction Detection — Conflicting beliefs are detected and resolved: demoted, archived as superseded, or kept as contextual variants.
- Hybrid Retrieval — Vector similarity plus knowledge-graph traversal up to 2 hops, ranked by relevance, recency, and confidence. Sub-10ms p95 recall.
- Multi-Subject Isolation — One agent serves thousands of isolated end-subjects via anchors, sessions, and shared canon memory, with per-subject GDPR purge.
- Conversation Extraction — Automatically extract and classify memories from transcripts. Multi-LLM: OpenAI, Anthropic Claude, Google Gemini, Cerebras.
Benchmark
LongMemEval (ICLR 2025): 91.4% overall accuracy with a reproducible public harness — knowledge update 100%, abstention 100%, multi-session 90.2%, temporal reasoning 82.3%.
Who Is Hakuya For?
- Agent & Copilot Builders — Anyone building AI agents, assistants, or copilots — in code or on an agentic / no-code platform — whose mistakes cost money.
- Regulated Industries — Health, finance, and legal teams that need auditable agent memory.
- Multi-Subject B2B — One support, clinical, or concierge agent serving thousands of isolated end-users.
- LLM App Builders — Persistent, trustworthy user context beyond a single session.
Integration & Tech Stack
REST API, Python SDK (pip install engram.to), LangChain integration (pip install langchain-engram), and an MCP server for Claude Desktop, Claude Code, Cursor, and Windsurf. Built with Go (single binary), PostgreSQL, pgvector, and Docker — self-hosted with one command.
Open Source
Hakuya is fully open source under the Apache-2.0 License. View on GitHub. LLM-readable context: llms.txt and llms-full.txt.