Fully-local long-term memory for AI agents built on two pillars: layered long-term memory (a semantic pyramid L0 Conversation → L1 Atom → L2 Scenario → L3 Persona) and symbolic short-term memory that offloads verbose tool logs to files while keeping a compact Mermaid 'canvas' in context. Distributed as a TypeScript/npm package; integrates with OpenClaw and the Hermes gateway.
- Storage
- Dual-layer heterogeneous storage: bottom layers (atomic facts, logs, raw tool outputs as refs/*.md, step summaries as jsonl) persist in a database for full-text retrieval; top layers (personas, scenes, Mermaid canvases) are stored as human-readable Markdown. A deterministic node_id drill-down path links each abstraction back to ground-truth evidence.
- Retrieval
- Hierarchical recall: the agent attends to the top-layer Persona/canvas in context and drills down to mid-layer indexes (Scenario/jsonl) and bottom-layer raw text (L0/refs) via node_id when detail is needed. Vector search over the stored layers.
- Self-host
- Self-host: moderate
- License
- MIT
- Pricing
- Open-source (MIT), free to self-host; fully local with zero external API dependencies. · Free / OSS
- GitHub stars
- 6,292
- Last release
- 2026-05-28
- Last commit
- 2026-06-26
- First catalogued
- 2026-06-28
Strengths
- Layered design (L0→L3 semantic pyramid) avoids flat vector dumps and preserves a traceable drill-down path back to source evidence
- Symbolic short-term memory (Mermaid canvas + history offloading) targets token reduction on long-horizon tasks
- Fully local with no external API dependency; backed by Tencent
Watch out
- Young project (created 2026-04) — the API and storage surface may still be moving
- License: README badge says MIT and the LICENSE file confirms MIT, though GitHub auto-classifies it 'NOASSERTION' (non-standard header)
- README carries self-reported benchmark numbers (WideSearch, SWE-bench, AA-LCR, PersonaMem) — record those via harvest-benchmarks with selfReported:true, not on this card
Best for
- Long-horizon agent tasks needing token-efficient, fully-local memory with traceable layered recall
Benchmark results
No sourced results yet.
Sources
- TencentDB Agent Memory README (vendor)
- GitHub API repo metadata (stars, release) (third-party)
- LICENSE (MIT, Copyright Tencent) (vendor)
Last verified 2026-06-28 · updated by discover-frameworks