StackMap
Subscribe

agentmemory vs deja-vu

Persistent memory for coding agents on the iii engine: MCP server with 53 tools, 12 auto-capture hooks, hybrid search + knowledge graph, zero external DBs. Claims 95% R@5 and 92% token cuts. — versus — Indexes the session histories your coding agents already wrote — 17 harnesses, months retroactive — and serves recall over MCP. 84.9% hit@1 on LongMemEval-S, no LLM, no embeddings. One Go binary.

The curated verdict

Both serve coding-agent memory over MCP. agentmemory is a 53-tool capture-everything suite; deja is one binary that only indexes and recalls — pick by how much machinery you want.

agentmemorydeja-vu
Stars26k556
Forks2.2k40
LanguageTypeScriptGo
LicenseApache-2.0MIT
Last activity2 days agotoday
Topicsmemory, codingmemory, coding
Curated connections32

agentmemory — the curator's take

The benchmark-forward entry in a crowded field — '#1 on real-world benchmarks' is self-run, so weigh it accordingly; what's independently real: 1,428 tests, the viral design gist it implements (Karpathy's LLM-wiki pattern plus confidence scoring and lifecycle), and the smoothest onboarding in the category — hand your agent one URL and it installs itself. The iii engine is your infrastructure bet: a server on :3111, not a library. A 53-tool MCP surface is the opposite of the plain-files philosophy — richer, but agents need guidance to use it well (they ship 15 skills for exactly that reason). Same prolific author as pro-workflow and tailclaude — expect fast movement, budget for churn.

deja-vu — the curator's take

The inversion is the insight: every memory tool starts empty and records forward; deja starts FULL from history already on disk, and the no-LLM/no-embedding design means ~1.5ms search, zero keys, zero cost. When NOT: it remembers what your agents did, not curated knowledge — there's no write path for distilled lessons; and its benchmark numbers are self-published, so the standing caution on this shelf applies: benchmark recall on your own corpus.