Niches · Agent memory

Remember a user's facts and preferences and bring the relevant ones back on later turns

Four or more companies' docs show a full fit.

5companies, full fit
14companies, fit with a limit
55companies carry the tag
22share of the category's work
9first-hand accounts

What counts as done

A search scoped to the user, in one call, with no other user's facts.

Measured, for our own tests: the fact from an early session is among the top 5 results, and boilerplate is not stored as memories (under 10% of stored entries are restatements of the system prompt). No docs we read state these figures, so they don't count against a listing here.

How agents ask for it: remember that I prefer window seats, what do you know about this user, pull in what we discussed last week before answering

Does the job 5

The docs show it meets the test above, and nothing we read says otherwise.

ListingWhy
Amazon Bedrock AgentCore Memory A
Amazon Web Services, Inc. (regional AWS entities by account location)
Namespace templates with {actorId} scope records per user, RetrieveMemoryRecords needs a namespace and returns records by relevance with score and createdAt, and consolidation avoids duplicates. Mission Cloud used namespaces per instructor in production. source
EverOS BB
Evermind AI, Inc.
add takes messages with sender_id, flush extracts episodes and a profile, and search filters by user_id, app_id and project_id returning episodes with timestamps and atomic facts. Markdown lands under users/<user_id>. source
Agent Platform Memory Bank B
Google LLC
Each scope such as {user_id} is an isolated collection, generation extracts and consolidates facts in the background, and similarity retrieval is scoped to one identity. source
Memori D
Memori Labs Inc.
entity_id, process_id and session_id scope every memory, augmentation extracts facts in the background and recall returns facts with id, content, similarity and date_created for the current entity. source
past.dev BB
Revo Inc.
A recall as an identity returns ranked documents with occurredAt, content and verbatim source excerpts, reaching only that identity's audiences; text is searchable within seconds of a completed send. source

Does the job, with a limit 16

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

ListingWhy
AI Memory B
Fabio Akita
Memory is a per-project wiki plus a cross-project profile and a reserved _global scope that memory_query returns beside project hits; there is no user id scope or fact extraction, so it suits one developer's preferences, not a per-user chat product. source
Basic Memory A
Basic Memory LLC
Notes are Markdown the agent writes with bm_write; there is no extraction from turns, and scope is a project rather than a user id, so per-user facts need one project per user and an agent that writes them. source
claude-mem C
Alex Newman
Observations are scoped by project name, not user id, and search filters by project and date; facts about a person land only where the agent or save_memory puts them. Suited to one developer, not a multi-user product. source
Cognee D
Topoteretes UG (haftungsbeschränkt)
Datasets are owned by a user and session cache is keyed by user and session, but the MCP default dataset is agent-scoped (for example cursor_vscode_memory), so per-user isolation needs API mode with one user per caller. source
Egoist AI Passport C
Egoist Machines, Inc.
recall serves only categories the owner has granted passes for; a bare recall with no passes returns guidance, and a sealed Passport needs the owner to unlock. The person's own agent gets their memory, after approval round trips. source
Graphiti D
Zep Software, Inc.
group_id namespaces hold one user's graph and search_memory_facts returns facts by hybrid search, but every add runs several LLM calls and a burst of turns hit provider rate limits with no record of what landed (issue 290). source
Hindsight D
Vectorize, Inc.
One bank per user with user:<id> tags, four parallel retrieval strategies and consolidated observations do the job, but one tool-heavy OpenClaw user burned 3.31M tokens in 30 minutes and found retainToolCalls needed a gateway restart (issue 1573). source
Honcho B
Plastic Labs, Inc.
Each user is a peer with a reasoned representation, peer card and chat endpoint, but reasoning runs in a background deriver, and a self-hosted user saw empty representations until small batches were flushed in 3.0.11 (issue 494). source
Hy Memory C
Tencent
add(user_id) and search(user_ids) return profile, normal and proactive channels over an eight-layer chat memory, but reconcile fell back to ADD and duplicated memories when store methods ignored source parameters (issue 26), and search returns every listed user's memories. source
LangMem D
LangChain (the licence file gives no legal form)
Namespaces such as ("memories", "{user_id}") set from config scope each user, and the store manager extracts facts in the background; but one user following the memory-agent example saw no records appear in Postgres with no reply (issue 154). source
Mem0 Platform + MCP E
Embedchain, Inc. (DBA Mem0)
Search with a user_id filter returns a user's facts in one call and the Platform pulls earlier turns as context; but a 32-day audit found 52.7 per cent of entries were system-prompt restatements (issue 4573), and free adds need your own LLM key after 30 days. source
MemMachine C
MemVerge, Inc.
Each project is an isolated namespace and the MCP defaults proj_id to mcp-{user_id}, with semantic profile memory and episodic search sorted by timestamp; but a search under a shared group returned another user's memory, which maintainers called normal (issue 320). source
MemPalace B
MemPalace (open-source project)
Verbatim drawers in wings and rooms are searched semantically with no LLM, so nothing is restated; but there is no per-user scope beyond wing naming, and a survey found contradiction detection absent from the code (issue 27). source
Supermemory API + MCP C
Supermemory
A namespace per user is a hard boundary, search and a profile endpoint return the user's facts, and status is polled to done; but one week of daily use stored 52 memories at 71 per cent accuracy with 16 false positives from over-extraction. source
TencentDB Agent Memory C
Tencent
Turns are distilled from L0 conversations to L1 atoms, L2 scenarios and L3 personas scoped by team, agent, user and task ids, but L1 search is keyword based, the bearer gate is off by default and headless runs need three headers or the proxy skips memory. source
Zep BB
Zep Software, Inc.
Each user has a graph by UUID and graph.get_context returns a context block with facts and their valid_at dates in one call; but the terms grant Zep a licence to train on customer data (listing weakness), so personal history cannot go in. source

Doesn't do this job 3

Listings in the category that do something else, with the reason.

ListingWhy
Context Mode Engine C
MKSF LTD
The Engine indexes tool output and the session log; the facts-with-two-clocks Memory is a feature of the paid hosted Gateway, a separate product. source
Graphicle BB
KarLex AI, Inc.
Memory is a project's reviewed decisions, procedures and notes anchored to code and documents, not per-user facts from conversation turns. source
Tokenrip B
Quintel, LLC
A shared workspace of files, tables, tasks and pinned decisions; no fact memory or semantic recall, and the vendor's FAQ says it is not a memory tool. source

What stops agents finishing

  • adds need an LLM key the free tier stops providing after 30 days (Mem0)
  • cloud plugin only on a paid plan (Supermemory OpenClaw at $19 a month)
  • vendor trains on free-tier data, so personal history can't go in

Around it

Before: chat.turn. After: llm.prompt-assembly.

Where this comes from

Weight 22 of 100 in Agent memory: 9 of 36 accounts; the default job in every vendor doc and the one the chat-product accounts (Mem0 HN thread, OpenClaw users) describe

The same as data: /api/v1/use-cases/agent-memory.user-facts-recall.json, or the MCP tool get_use_case with id agent-memory.user-facts-recall. Other jobs in this category: /api/v1/niches?category=agent-memory.

Draft, read 2026-11-12 from public docs and first-hand accounts, not yet reviewed by a person. A mark says what the docs show, not what a run measured.

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.