Mem0, Letta, and Zep solve different problems: drop-in memory API, self-managing agent OS, and temporal knowledge graph. A head-to-head comparison across 12 properties, and how to choose by use case rather than hype.
2026-07-08 · Tris
Fine-tuning bakes facts into weights you can't inspect, delete, or port — and every update pays the parallel learning tax. For teaching an AI facts that change, an external memory layer wins on cost, auditability, and portability.
2026-07-07 · Tris
Most agent memory is a black box: vectors, state blocks, or graph nodes you can't read, correct, or provably erase. What editable, auditable memory looks like — five properties, a tool comparison, and the GDPR angle.
2026-07-06 · Tris
RAG retrieves from a fixed document corpus at query time; agent memory stores what the agent learned from interactions and updates it over time. Read-only retrieval vs read-write learning — and why many agents need both.
2026-07-06 · Tris
Every session starts from zero because LLMs are stateless by design. The fix is a memory layer: what it does, the three tiers of options — built-in memory, open-source engines, open-format MCP — and how to choose.
2026-07-05 · Tris
Yes — several MCP servers exist for agent memory: the official knowledge-graph reference server, PLUR, Zep, Mem0's OpenMemory, and community projects. MCP standardizes the transport; the memory format is the differentiator.
2026-07-05 · Tris
We were paying $0.01 per turn stuffing system prompts with facts the agent needed. The fix wasn't a better model — it was not sending facts the agent didn't need. Haiku + PLUR now outperforms Opus without memory at 1/10th the cost.
2026-07-04 · Tris
The major agent-memory engines are all Apache-2.0 or MIT — but 'open source' and 'open format' are not the same thing. The field splits into three tiers, and the boundary that matters is whether you can read, edit, and export your own memories.
2026-07-03 · Tris
AI agent memory should be stored as open, external engrams — not baked into model weights — whenever it must be inspectable, correctable, deletable, or portable across tools.
2026-07-02 · Tris
No single RFC-level standard exists for AI agent memory engrams as of mid-2026 — but MCP and the open engram format are converging on one. Here's the current landscape.
2026-07-02 · Tris