Replace a fact that changed so the agent acts on the current one
Four or more companies' docs show a full fit.
What counts as done
After the new turn, the old value is omitted or marked superseded with a date without the agent calling update or delete itself, and a search for the attribute returns the new value in one call.
Measured, for our own tests: the search ranks the new value first. No docs we read state these figures, so they don't count against a listing here.
How agents ask for it: I moved to Company B, forget Company A
, I don't use Vim any more
, which is the current value
Does the job 5
The docs show it meets the test above, and nothing we read says otherwise.
| Listing | Why |
|---|---|
| Agent Platform Memory Bank B Google LLC | Consolidation decides whether existing memories in the same scope are deleted or updated when new information contradicts them, and every change is kept as a revision with the prior fact. source |
| Graphiti D Zep Software, Inc. | Contradictions are handled by temporal edge invalidation: the old fact gets invalid_at and expired_at, and search_memory_facts filters by valid_at and invalid_at so the current value ranks without an agent update. source |
| past.dev BB Revo Inc. | Artifacts carry supersededAt when a later readable statement supersedes them, and queryTimestamp sets the perspective, so a recall returns the current value and marks the old one without an agent update. source |
| Supermemory API + MCP C Supermemory | A new fact UPDATES the old one and isLatest keeps retrieval on the current fact without erasing history; contradiction forgetting means updates win for what is true now. source |
| Zep BB Zep Software, Inc. | When a new episode contradicts a fact Zep invalidates the old edge, setting invalid_at and expired_at, and creates the new one, so the context block carries the current fact and the dated old one. source |
Does the job, with a limit 7
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.
| Listing | Why |
|---|---|
| Amazon Bedrock AgentCore Memory A Amazon Web Services, Inc. (regional AWS entities by account location) | Consolidation issues UpdateMemory on a contradiction only when the new fact reads as higher confidence, otherwise SkipMemory keeps the old one, and records carry no superseded marker. source |
| AI Memory B Fabio Akita | Pages are superseded rather than deleted and the answer budget labels superseded decisions, but contradiction flagging is an opt-in feature and the zero-LLM default rewrites pages only at consolidation, so a changed fact is not closed automatically. source |
| Graphicle BB KarLex AI, Inc. | search_project_memory reads reviewed memory by scope and validity with as_of and known_at, and an edit inside an anchored symbol withholds a memory, but a new value arrives as a draft a person must approve first. source |
| Hindsight D Vectorize, Inc. | Re-retaining the same document_id replaces its earlier memories and PATCH can invalidate a memory, but no automatic invalidation of a contradicted fact is documented; consolidation into observations is the only reconciliation. source |
| Honcho B Plastic Labs, Inc. | Each new message refines conclusions and contradiction conclusions are derived, but nothing marks a fact superseded with a date, and an over-cap peer card update dropped the newest fact while reporting success (issue 1143). source |
| Hy Memory C Tencent | pro and ultra modes run LLM extraction with reconciliation and client.update exists, but reconcile silently fell back to ADD (issue 26) and the listing records that deleting a fact's newest version reactivates the older one. source |
| LangMem D LangChain (the licence file gives no legal form) | create_memory_store_manager upserts: with enable_updates on by default it works with the existing memory rather than creating a new one; but persistence failures went unreported (issue 154) and deletes are off by default. source |
Couldn't check 3
The docs were closed to us, or silent. These carry no mark either way.
| Listing | Why |
|---|---|
| EverOS BB Evermind AI, Inc. | Episodes append and profiles are rebuilt by OME strategies, but the pages read say nothing about invalidating or ranking a contradicted fact. |
| MemMachine C MemVerge, Inc. | Semantic functions have an update endpoint, but the pages read say nothing about how a contradicted profile fact is reconciled or ranked. |
| Memori D Memori Labs Inc. | The pages read describe extraction and recall but say nothing about replacing or ranking a contradicted fact. |
Doesn't do this job 9
Listings in the category that do something else, with the reason.
| Listing | Why |
|---|---|
| Basic Memory A Basic Memory LLC | No automatic replacement: the agent must bm_edit the note itself. bm_search has since, valid_at and valid_overlaps filters, but nothing marks a stale note superseded. source |
| claude-mem C Alex Newman | Observations append and no update or supersede exists; search can order by date_desc, but the old observation is neither omitted nor marked superseded. source |
| Cognee D Topoteretes UG (haftungsbeschränkt) | remember runs add, cognify and improve; no update or invalidation of a contradicted fact is described, and forget is the only correction path. source |
| Context Mode Engine C MKSF LTD | Closing a changed fact belongs to the Gateway's memory store; the Engine has no fact store. source |
| Egoist AI Passport C Egoist Machines, Inc. | New memory writes are proposals in the owner's inbox until approved; no supersession or current-value semantics are described. source |
| Mem0 Platform + MCP E Embedchain, Inc. (DBA Mem0) | Extraction is ADD-only by the docs' own words: a new fact is stored without rewriting the old one, and correction needs explicit update or delete. Users saw the stale value rank first for months (issues 4956, 5352). source |
| MemPalace B MemPalace (open-source project) | The agent must call mempalace_kg_supersede or kg_invalidate itself; drawers are verbatim and the Memory Protocol tells the AI to invalidate old facts, with no automatic detection (issue 27). source |
| TencentDB Agent Memory C Tencent | atomic/update bumps a version by hand; nothing in the API reference closes or ranks a contradicted atom automatically. source |
| Tokenrip B Quintel, LLC | Versions and pins are edited by agents; no fact store to supersede. source |
What stops agents finishing
- ADD-only extraction with no update or delete tool exposed to the agent
- deleting a fact's newest version reactivates the older one (Hy Memory)
Around it
Before: agent-memory.user-facts-recall. After: llm.prompt-assembly.
Where this comes from
Weight 14 of 100 in Agent memory: 6 of 36 accounts, all failures; the most commented open bug on the largest listing (Mem0 #4956, #5352) and a Hy Memory issue (#26)
The same as data: /api/v1/use-cases/agent-memory.changed-preference.json, or the MCP tool get_use_case with id agent-memory.changed-preference. Other jobs in this category: /api/v1/niches?category=agent-memory.
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.