Open-source agent memory separates third-party claims from user facts
Deep-Thinker-01 · reddit · 2026-07-23
Open-source agent memory with typed third-party claims and evidence-grounded abstention
The post introduces an open-source memory system for agents that keeps user facts and third-party claims separate instead of flattening everything into one store. It is designed around three recurring failure modes from the research: poisoning, confident-when-wrong behavior, and staleness.
How it works
- Uses a typed graph plus dated episodes as the store of record.
- Stores external information as thirdpartyclaim edges rather than treating it like first-party memory.
- Uses derivedfrom to cap trust when provenance is mixed, such as when the agent quotes an event from someone else’s text.
- Adds an evidence-grounded abstention gate, so the agent can refuse or hedge when evidence is insufficient.
- Supports functional supersession with history so newer facts can replace older ones without deleting provenance.
Practical details
- Includes a remember → recall partition → answer workflow.
- Exposes Store / Complete seams, an MCP server, and JSONL export/import.
- The author also links self-checks, docs, and the repo, emphasizing that the design avoids lock-in.
Related event: Developers Debate the Real Challenges of AI Agent Memory(6 posts)→
More from Research
- Artificial Analysis breaks down all sub-evaluations in Intelligence Index v4.3.2 — ArtificialAnlys · 2026-10-27
- Infini-AI Lab open-sources Hermit, a framework for agent harness evolution and continual learning — BeidiChen · 2026-10-27
- Meta FAIR's Self-Direction turns agent failures into reusable LoRA skills, +13.8pp on Terminal-Bench — karen_ullrich · 2026-10-27
- DONIC hits 86.79% on bimanual dexterous tracking, crushing prior SOTA's 9.43% — liuziwei7 · 2026-10-27
- Why don't more LLMs ship 10M-token context? A researcher asks about attention tradeoffs — agarwl_ · 2026-10-27
- PSNR below 15 dB in papers likely means the metric was misused, researcher warns — kwangmoo_yi · 2026-10-27