The related-systems reviews are the best proof the knowledge layer works

Type: kb/types/note.md · Status: seedling

We have reviews of 15 agent memory systems — Mem0, Graphiti, Cognee, Letta, A-MEM, AgeMem, Ars Contexta, Thalo, ClawVault, CrewAI Memory, Siftly, sift-kg, Spacebot, Decapod, Hindsight, SAGE, and getsentry/skills — plus a comparative review that synthesizes across all of them along six architectural dimensions.

These reviews were mostly vibed — minimal human input, agent-driven analysis. And they're already useful. Someone choosing an agentic memory system today would benefit from:

  • The agency model taxonomy (who decides what to remember — agent-self-managed vs developer-managed vs human-collaborative vs RL-trained)
  • Per-system reviews that go beyond README claims to actual code analysis
  • Cross-cutting patterns: what converges (progressive disclosure, extraction automation) and what diverges (storage model, link structure, curation operations)
  • Honest gap analysis — including gaps in our own system

Why this matters for positioning

  1. It's a live demo. The reviews themselves were produced using the KB. Each new review built on the previous ones — the comparative framework emerged from accumulation, not upfront design. This is the compounding effect in action.

  2. It solves a real problem people have right now. The agent memory space is exploding and nobody has a good map. Our reviews provide one. This is immediately useful regardless of whether someone adopts our framework.

  3. It shows the difference between retrieval and composition. A RAG system could surface individual facts about Mem0 or Graphiti. It could not produce the comparative review — that required traversing links across reviews, recognizing patterns, and synthesizing a novel argument (the agency trilemma). That's what composable knowledge enables.

  4. Mostly vibed = low barrier. The reviews demonstrate that the framework doesn't require heroic human effort. An agent with structured knowledge and good conventions produces useful analysis with minimal steering.

How to use this in positioning

  • Lead with the comparative review as a flagship artifact — it's the most immediately valuable thing in the KB for an external audience
  • Individual system reviews serve as proof of depth — pick 2-3 that the audience would recognize
  • The "mostly vibed" story addresses the "sounds like a lot of work" objection
  • Frame as: "this is what your agents could produce about YOUR domain if they had a knowledge layer"

Open Questions

  • Should we publish the reviews standalone (blog post, GitHub pages)?
  • Do we need to add caveats about the vibed nature, or is transparency about it actually a strength?
  • Which reviews need a quality pass before external exposure?