How to read this

A landscape catalog of AI memory systems — products, frameworks, papers, and theoretical proposals — built as the substrate for cross-system comparison, adoption signal, and trend analysis.

What this is

523 systems across 22+ sections, scored on 67 columns (memory model, claims, performance, deployment, governance, …), connected by 247 typed edges (built-on, cites, succeeds, …).

Numbers are hardcoded from landscape.json at last commit; the true count is whatever the table view shows on first load.

Tier badges

Every record carries an ordinal tier from T1 (most battle-tested) to T5 (most speculative). Tier is not a quality score — a T5 manifesto and a T1 commercial product are different artefacts, both worth tracking.

  • T1 Battle-tested. Real customers, measurable enterprise adoption (commercial + mature OSS).
  • T2 Established / mature OSS. Significant traction, clear maintainer team, in production at known users.
  • T3 Peer-reviewed. Published at a recognised venue (ACL, EMNLP, NeurIPS, ICLR, …) with code.
  • T4 Preprint / unpublished. arXiv (or similar) with implementation but not yet peer-reviewed.
  • T5 Theoretical or informal. Idea-stage; blog post, manifesto, taxonomy entry, no code.

The 7 taxonomy axes

Each record is placed on seven independent axes. A record can have multiple values per axis (e.g. Mem0 is vector + graph + kv on the storage axis); exactly one value per axis is marked primary.

Storage
Where memory data sits — vector / graph / kv / file / latent / hybrid / …
Retrieval
How reads happen — semantic / structural / hybrid / parametric / …
Persistence
Durability of the memory — session / process / persistent / …
Update
How memory gets written — explicit / implicit / batch / …
Unit
Granularity of what's stored — fact / chunk / event / trajectory / …
Governance
Who decides what gets remembered — auto / curated / mixed / …
Conflict
How contradictions resolve — latest / vote / merge / ignore / …

Cell status indicators

Every cell in the 67-column table carries a status so a reader can tell "data" from "we tried and found nothing" from "doesn't apply here".

  • real-data Actual data, sourced and (usually) cited.
  • not-applicable Column doesn't apply to this record (e.g. funding on a research paper).
  • depth-floor-reached Searched in good faith; no info found. Honest gap.
  • no-data Placeholder for an un-researched cell. Rare in terminal data.

How edges work

The Graph and Lineages views are powered by a separate file of typed directed edges. Multiple edges between the same pair are allowed if the type differs (Letta succeeds MemGPT and is by the same team as MemGPT — both useful).

  • built-on — source's product is implemented on top of target (target is a dependency or runtime).
  • extends — source extends or generalises target's method, keeping the core idea.
  • forks — source is a literal code fork of target.
  • integrates-with — source has a first-class integration / connector / adapter to target.
  • competes-with — positioned by the market (or themselves) as alternatives in the same buyer's mind.
  • inspired-by — source cites target as conceptual inspiration without building on or extending it.
  • cites — source's paper cites target's paper; populated from Semantic Scholar.
  • same-team-as — same authors, lab, or company across two systems.
  • succeeds — explicit successor / next-version by the same team (e.g. MemGPT → Letta).

Methodology & provenance

Every entry was sourced from one of:

  • Curated lists (Agent-Memory-Paper-List, Awesome-GraphMemory).
  • Survey papers (arxiv.org/abs/2512.13564, arxiv.org/abs/2508.10824).
  • Benchmark leaderboards (LongMemEval, LoCoMo, ConvoMem).
  • Vendor websites and academic venue pages.
  • Targeted research-agent sweeps (~25 agents over multiple rounds).

Coverage of the memory-shaped core: ~88–92% as of the last terminal pass. URLs were verified at time of entry. Claims (the right-hand columns) are vendor-stated unless otherwise marked.

Scope. The catalog covers memory systems for AI and the adjacent infrastructure that touches them — training platforms and dataset stores, generic vector / search systems, and agent frameworks regardless of whether they ship a first-party memory layer. The goal is comprehensive landscape coverage of the sphere, not a strict memory-only filter. Coverage of the adjacent categories is lower than the memory core and being expanded incrementally.

Acknowledgements & complementary resources

  • dbdb.io — Carnegie Mellon's database-of-databases. Same form, different domain; roughly 17 systems overlap with this catalog.
  • Agent-Memory-Paper-List — source curated list for the research-side entries.
  • Awesome-GraphMemory — source curated list for graph-memory systems.
  • GitHub repo — issue tracker, build plan, decision log, raw landscape.json.