Agent Memory for Team Knowledge — Competitive Landscape & Lessons for AsDecided

Status: research document, not a RAC artifact. Per ADR-010 ("Documents Are Not Artifacts") and ADR-024 ("RAC Is Not a Content Store"), this is reference / working material, not part of the validated decisions/ corpus. Its actionable conclusions are captured as corpus artifacts — see "What this produced" below.

Scope: products and projects positioned as "agent memory for a team's knowledge" — systems that turn docs, chats, tickets, code, or decisions into a queryable store an AI agent reads — and where AsDecided (RAC) sits among them. Method: two multi-angle web-research passes with adversarial verification (3-vote, 2/3 to kill) and evidence-vs-marketing tagging; a second pass closed the coverage gaps (Cognee, Glean, Dust, native assistant memory) the first left open. Confidence and primary sources cited inline. Date: 2026-06.

Executive summary

The space splits on one axis, and AsDecided sits cleanly on the under-served side: LLM-distilled mutable stores vs human-reviewed versioned knowledge. The overwhelming majority — Mem0, Zep/Graphiti, Cognee, Letta, OB1, and the native memory in ChatGPT, Claude, Cursor — use an LLM to extract facts at ingest into a mutable database, vector index, or knowledge graph, and are per-agent or per-user. None routes knowledge through human review before it becomes authoritative. That is the right design for personal recall and the wrong design for a shared team source of truth, where a confidently-distilled stale "fact" is the exact failure mode the category cannot police.

Only a thin git-native, human-reviewed band overlaps AsDecided's positioning, and it has just two real occupants — Mainline and Kage — both of which appeared recently and neither of which governs a typed requirements/decisions corpus with ratified supersession. The convenient "only git-native team-knowledge engine" line is therefore no longer true and must be retired.

One-sentence positioning that the research supports:

AsDecided is the git-native, human-ratified source of truth for a team's typed decisions and requirements — knowledge that is authored and accepted by people, not distilled by a model — winning on the trust axis (deterministic supersession and provenance) that every LLM-distilled memory store structurally cannot own.

The axis

	LLM-distilled mutable store	Human-reviewed versioned knowledge (AsDecided)
How knowledge is made	LLM extracts facts at ingest	authored + ratified by a human
Source of truth	database / vector / graph store	files in git, `main`
Determinism	non-deterministic (model-dependent)	deterministic, reproducible
Supersession	model-judged or in-place overwrite	explicit, typed, history retained
Scope	mostly per-agent / per-user	shared team corpus
Review gate	none (agent writes autonomously)	pull-request review (ADR-065)

Category findings

1. LLM-distilled mutable stores

Mem0, Zep/Graphiti, Cognee, Letta, OpenBrain/OB1, Collaborative Memory.

Zep / Graphiti is the most sophisticated foil: an LLM extracts and resolves entities/relationships at ingest into a temporally-aware knowledge graph (Neo4j default; also FalkorDB, Neptune), with a bi-temporal model that invalidates superseded facts (sets t_invalid) rather than deleting them, and avoids LLMs at query time (~300ms P95). Storage is a graph DB, not git. [high] (arXiv 2501.13956, Graphiti)
Mem0 uses LLM-driven extraction into a hybrid vector+graph+KV store and updates facts in place via an LLM-decided ADD/UPDATE/DELETE/NOOP — no temporal validity windows. Scoped per user_id/agent_id/session: episodic memory, not a shared source of truth. [high] (arXiv 2504.19413, repo)
Cognee is an open-source (Apache-2.0) "AI memory platform for agents." Its Extract-Cognify-Load pipeline uses an LLM (Instructor-powered structured output) to build a typed knowledge graph at ingest, stored in graph + vector + metadata DBs (Kuzu / LanceDB / SQLite) — non-deterministic, no human-review workflow on agent-added knowledge. Self-hostable; commercial cloud tier exists. [high] (repo, grounding AI memory)
Letta Context Repositories are genuinely git-backed (every memory change auto-committed; agents clone the repo; subagents merge via worktrees) — but memory is per-agent and managed by LLM "sleep-time" reflection subagents, with no human-ratification gate. Architecturally close to AsDecided's medium, opposite on trust. [high] (Letta)
OB1 (OpenBrain) uses LLM "schema-aware routing" to distribute text into a PostgreSQL/pgvector store (Supabase or self-hosted); it does track provenance / derivation chains, but the source of truth is a DB, not git. [high] (repo)
Collaborative Memory (Accenture research, not a product) is notable for two AsDecided-relevant ideas in an otherwise LLM-distilled mutable store: private + shared tiers with access control, and immutable provenance metadata per fragment (creation time, contributing user/agents, resources accessed). [high] (arXiv 2505.18279)

Note for AsDecided: even the best supersession in this camp (Graphiti's bi-temporal invalidation) is model-judged at ingest. AsDecided's supersession is human-ratified and typed — the difference between "the model decided this fact is stale" and "a reviewer accepted that this decision supersedes that one."

2. Git-native human-reviewed knowledge — AsDecided's real neighbours

Mainline, Kage; Letta Context Repositories (per-agent, see above); Semiont (positioning-adjacent).

Mainline is the closest paradigm match: deterministic, git-native, no DB / no vectors / no LLM distillation / no embeddings. It stores agent intent and decisions as Git refs and notes alongside code, collaborates via fetch/branch/ merge, and binds decisions to the commits that produced them (commit-level provenance). [high] (mainline.sh, repo)
Kage is the closest match to AsDecided's positioning: git-native ("no account, no API key, no database — just files in git"), reviewed in the same pull request as the code, with strong provenance/staleness handling — citations validated against the repo at capture (hallucinated citations refused), stale memory withheld when its cited file is deleted, and kage pr check warns when a change invalidates team knowledge before the PR lands. [high, single vendor source] (kage-core.com)
Semiont (AI Alliance) targets the trusted/provenance/human-governed framing ("AI proposes, domain experts review"; W3C Web Annotation provenance) — but it uses an LLM to propose, and verification refuted its git-source-of-truth and direct-"agentic-memory" claims. Treat as positioning-adjacent, not git-native. [medium] (repo)

Note for AsDecided: Mainline and Kage capture agent intent / lessons; AsDecided governs a typed requirements/decisions corpus with ratified supersession and a validated graph (ADR-049, ADR-061, ADR-074). That is the narrower, defensible distinction now that "only git-native" is gone.

3. Enterprise RAG / search assistants — adjacent, not competitors

Glean, Dust.

Glean is enterprise "Work AI": a permissions-aware knowledge graph + vector store over indexed company content, with LLM-distilled retrieval. A search / knowledge-discovery product, not a versioned source of truth. [high] (Glean KG)
Dust is RAG at query time: semantic search over connected data sources feeds retrieved docs + the question into an LLM. No ingest-time fact distillation, no versioned corpus — an agent platform over company data. [high] (Dust RAG)

4. Native assistant memory — per-user, mostly distilled, no shared gate

ChatGPT, Claude, Cursor, Continue.

ChatGPT memory builds a persistent, LLM-distilled user profile. Claude memory is LLM-distilled for Teams/Enterprise (a wiki-style doc auto-updated on a cycle, no human-ratification gate); consumer Claude instead retrieves over raw conversation history with no AI-generated summaries. All per-user. [high] (Claude memory)
Cursor / Continue rules (.cursor/rules/*.mdc, .cursorrules) are the one native feature sharing AsDecided's medium: human-authored, git-versioned, team-shared files reviewed in the normal PR flow. But they are free-form instructions, not a typed, validated, supersession-aware corpus. [high] (Cursor rules)

Note for AsDecided: Cursor/Continue rules prove the substrate (git + PR review) is an accepted industry pattern, not a compromise — but they stop at free-form prose. The typed artifact model + deterministic validation is the layer above them.

Cross-cutting evidence

No agent-memory system ships a human-review promotion gate. A direct survey finding: "no system natively implements [a human-review step]; all assume the agent has authority to update memory directly." This is the gap AsDecided's two-gate capture model (ADR-077) and PR trust boundary (ADR-065) fill. [high]
Deterministic supersession beats LLM-judgment — measured. A 2026 preprint reports that replacing LLM-judgment with a candidate-extraction + max(serial) pipeline for memory freshness/conflict resolution yields +10.8 points on the FC-SH benchmark (67.2 → 78.0), and that LLMs degrade at tracking which knowledge version is current — external support for AsDecided's deterministic stance (ADR-066, ADR-080). [medium — recent preprint, cite as supporting evidence, not settled] (arXiv 2606.01435)

Where each player lands (verdict)

Player	Camp	Overlap with AsDecided
Zep / Graphiti	distilled mutable (graph)	adjacent — strongest supersession, but model-judged, DB-backed
Mem0	distilled mutable (hybrid)	adjacent — per-user episodic
Cognee	distilled mutable (graph+vector)	adjacent — no human gate
Letta Context Repos	git-backed, LLM-managed	adjacent — per-agent, no human gate
OB1	distilled mutable (Postgres)	adjacent — DB source of truth
Mainline	git-native, human-reviewed	direct paradigm overlap — agent intent, not typed corpus
Kage	git-native, human-reviewed	direct positioning overlap — agent lessons, not typed corpus
Semiont	LLM-assisted, human-governed	positioning-adjacent (git claim refuted)
Glean / Dust	enterprise RAG	adjacent — retrieval, not a corpus
ChatGPT / Claude memory	distilled, per-user	adjacent — personal memory
Cursor / Continue rules	git-native, human-authored	closest medium — free-form, not typed

Honest caveats & open questions

The two surviving differentiators are narrow but real: (a) no LLM distillation of judgment, and (b) a human-ratification gate over a typed corpus with explicit supersession. Mainline and Kage do PR review of agent intent/lessons; none ratifies typed decision/requirement artifacts the way AsDecided does.
"Only git-native" is retired. Mainline and Kage occupy the same substrate; the defensible claim is human-ratified + typed, not git-native exclusivity.
Refuted claims (excluded): Kage having an explicit supersession command with a newest-wins retain-both policy (1-2); Semiont being git-backed (0-3) or a direct "agentic memory" overlap (1-2); Cognee's temporal handling being merely an optional Graphiti add-on (0-3). Do not repeat these.
Vendor sourcing: Mainline, Kage, and Semiont positioning rests substantially on self-description (appropriate for what a product claims, not independently audited). The durable third-party facts are the Graphiti/Mem0 papers, the deterministic-supersession preprint, and Cursor's own rules docs.
Fast-moving space: most sources are 2025-2026; Mainline, Kage, Mem0 V3, Letta Context Repositories, and OB1 are all recent. Treat named feature deltas as perishable; lead with the durable trust-model distinction.
Tooling note: the second research pass's auto-synthesis step failed (returned a placeholder); these findings were recovered from the verified-claim layer beneath it (24 confirmed, 1 killed), not from a broken top-level summary.

What this produced (captured as corpus)

The competitive positioning conclusions were distilled into a growth-positioning requirement (this document is the reference behind it):

decisions/requirements/rac-growth-agent-memory-positioning.md (RAC-KVWHTV28J65S) — relate AsDecided to the agent-memory category on a docs/ comparison page (deliberately not the corpus README); name Mainline and Kage; retire "only git-native"; state the human-ratified + typed differentiator; and acknowledge the adjacents (Glean, Dust, ChatGPT, Claude, Cursor, Continue) honestly.

It complements the broader team-scale-landscape.md in this folder, which covers the adjacent decision/spec/portal/wiki categories and the freshness-gap thesis.