memory-layers-for-agents
ai-labs/studies/memory-layers-for-agents/context-v/
20 files here
profiles/19
Beads — Profile
Not one embedding anywhere in the recall path — retrieval is a SQL view and a graph walk over the agent's own work graph.
ByteRover CLI — Profile
The memory store is a directory of Markdown you can cat, grep, and diff, with its own git history. Recall is BM25, not embeddings.
Delta-Mem — Profile
Memory as a dense rank-8 state matrix inside each attention head, updated token by token — the model never looks anything up.
Graphify — Profile
The whole memory layer is one in-process NetworkX graph and a `graph.json` you can copy around — no database anywhere.
Graphiti — Profile
`expired_at` is when the engine retired a fact; `invalid_at` is when it stopped being true in the world. Routinely different dates.
Hindsight — Profile
`fact_type` is a database CHECK constraint: world facts and first-person experiences travel through different extraction prompts.
Honcho — Profile
Collections are unique on (observer, observed, workspace) — Honcho indexes memory by relationship rather than by user.
Letta — Profile
The one entry where the agent itself — not an extraction pipeline, not a background summarizer — is the memory manager.
Mem0 — Profile
Despite the graph-memory marketing, the graph is an array of linked memory IDs hanging off spaCy-extracted entities.
MemPalace — Profile
The counter-bet: extract nothing. Raw 800-character chunks score 96.6% R@5 on LongMemEval at zero LLM calls.
Neo — Profile
Familiar high-confidence queries get dialled down to low reasoning effort — memory shows up as a smaller bill, not better accuracy.
OpenViking — Profile
Memories, resources, and skills collapse into one `viking://` tree, and retrieval is a recursive directory walk, not a vector query.
Paxm — Profile
Paxm is not a memory store — it is routing and durability in front of seven-plus pluggable backends, and it does zero semantic dedup.
RetainDB — Profile
Supersession as an actual indexed column — `validUntil`, `supersededBy` — rather than something the ranker is hoped to handle.
StateBench — Profile
The trade nobody advertises: more context raises decision accuracy and superseded-fact resurrection together.
Supermemory — Profile
The most polished client surface in the study, wrapped around an engine that is not in the repo — every SDK points at a hosted API.
Understand-Anything — Profile
When an LLM authors your graph, the schema layer stops being a validator and becomes a repair shop — alias maps fold synonyms back.
Volt — Profile
Numeric thresholds decide when to compact, and the model is only ever asked to write summaries. Long-horizon coding as a DB problem.
memory-layers-for-agents/profiles