You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Consider promoting basic-memory-skills lifecycle/defrag conventions (stale-fact marking) into core #1588
Thanks again for the great work on Basic Memory — following up on a separate point from my earlier note (#1372/#1587) about progress visibility.
I noticed the core note format (frontmatter + free-text Observations with optional [category] tags) has no built-in way to distinguish a durable fact from a point-in-time status, or to mark a note/fact as superseded or expired. In practice, older notes can end up contradicting newer ones, and nothing in the core schema or retrieval flags that.
I see the companion repo basic-memory-skills already addresses this with Memory Defrag (removing stale info, merging conflicts), Memory Reflect (consolidating recent notes into long-term memory), and Memory Lifecycle (folder-based status/archiving, "archive never delete"). That's a solid approach, but it currently lives as an optional workflow layer rather than something the core tool is aware of.
Would it be worth considering a lightweight, optional signal in core, for example a status: or as_of: frontmatter field, or a freshness/recency weighting at search time? That would mean, first, users without the extra skills installed still get some protection against stale or contradicting notes surfacing in search, and second, the lifecycle/defrag conventions in basic-memory-skills would have an underlying primitive to build on, rather than relying purely on file organization.
Happy to share more detail on the specific workflow that surfaced this if useful. Thanks for considering it.
@Flynn-Kawolsky I’d be interested in the concrete workflow you offered. I’m studying correction and review of agent memory while building GFT Map.
If this is still occurring, could you share one non-sensitive example of the older observation, its replacement, and what a later search returned? Does status filtering or archiving already resolve it, or what do you still need to check or change yourself before continuing?
Compare with the same call without status. This is exact, whole-note filtering: unlabelled notes are excluded too. It still requires someone to maintain the labels, and it doesn't retire an old observation inside a still-active note. Ordinary searches don't apply this filter automatically.
I've checked the released source, not run this against your notes. If you try it, the useful distinction would be whether whole-note exclusion covers your case, or whether the remaining work is deciding which individual observations are still valid. No real note content needs to be shared.
Adding a data point to this request. I keep a status field in frontmatter (for example active, inactive or superseded). In my own retrieval eval, notes marked superseded or inactive still take a noticeable share of the top-5 results for questions whose answer is a current note, even though they rarely rank first. The 1.0 metadata filters cover most of this, but every caller has to remember to pass the filter. Also, $in drops notes that have no status at all, so I can't ask for "status is not superseded" and still keep notes without a status. Would a config setting for a default search filter (with a per-call way to include historical notes), plus a $nin or $ne operator, be a better fit than time decay (#603)? If #1973 lands, schema validation could then check that the field is present.
For the proposed default filter, would a small fixture covering active, superseded, explicit null, and a missing status key help settle the contract? In particular, should $ne: superseded retain both null and missing values?
I'd also test a limit of 5 with superseded notes ranked ahead of at least five eligible notes, and check that the filtered call still returns five eligible results. That would catch filtering only after the candidate limit has been applied. The per-call historical override could then be checked against the same fixture.
I haven't run this against Basic Memory; these are suggested acceptance tests, not a bug report.
I hit this running a company's memory on Basic Memory. What worked for me was fixing it at write time: one fact per observation line, each with a since date and a source. When a value changes, the old line stays, tagged #replaced with a pointer to the new one, so each property has exactly one #current line.
On your top-5 point: this doesn't stop old lines from ranking. Replaced lines still come back in search (I search with items: ["observation"]). What changes is that every hit says whether it's current, so the model answering can skip the stale ones. I haven't tested filtering on the tag itself.
On LongMemEval's 78 knowledge-update questions (Haiku answering, Sonnet judging), the whole brain in the prompt scored 69/78 vs 63/78 for reading the full history. After a fix to the writing rules I re-ran 71 of those questions, answering only from BM's top 15 hybrid search hits (top 15, not your top 5): 64/71, against 65/71 with the whole brain in the prompt. Self-run, one answer model, one judge, so take it as a signal.
The writing rules are a free Claude skill: https://github.com/topc-social/brainy. @Flynn-Kawolsky, since you already have a retrieval eval, would you try it and tell me if stale facts stop reaching the answer? Even a "no change" would help.
Activity
CoralLips commentedon Oct 3, 2026
@Flynn-Kawolsky I’d be interested in the concrete workflow you offered. I’m studying correction and review of agent memory while building GFT Map.
If this is still occurring, could you share one non-sensitive example of the older observation, its replacement, and what a later search returned? Does status filtering or archiving already resolve it, or what do you still need to check or change yourself before continuing?
CoralLips commentedon Oct 6, 2026
@Flynn-Kawolsky One concrete baseline for this proposal: the released v0.23.2 search tool already accepts a frontmatter
statusfilter.A small check on two disposable note copies:
status: supersededand its replacementstatus: activein their existing YAML frontmatter, then let Basic Memory sync them.{ "query": "shared phrase", "project": "your-project", "search_type": "text", "entity_types": ["entity"], "status": "active" }Compare with the same call without
status. This is exact, whole-note filtering: unlabelled notes are excluded too. It still requires someone to maintain the labels, and it doesn't retire an old observation inside a still-active note. Ordinary searches don't apply this filter automatically.I've checked the released source, not run this against your notes. If you try it, the useful distinction would be whether whole-note exclusion covers your case, or whether the remaining work is deciding which individual observations are still valid. No real note content needs to be shared.
Flynn-Kawolsky commentedon Oct 30, 2026
Adding a data point to this request. I keep a status field in frontmatter (for example active, inactive or superseded). In my own retrieval eval, notes marked superseded or inactive still take a noticeable share of the top-5 results for questions whose answer is a current note, even though they rarely rank first. The 1.0 metadata filters cover most of this, but every caller has to remember to pass the filter. Also,
$indrops notes that have no status at all, so I can't ask for "status is not superseded" and still keep notes without a status. Would a config setting for a default search filter (with a per-call way to include historical notes), plus a$ninor$neoperator, be a better fit than time decay (#603)? If #1973 lands, schema validation could then check that the field is present.OakcoderX commentedon Nov 7, 2026
For the proposed default filter, would a small fixture covering
active,superseded, explicitnull, and a missingstatuskey help settle the contract? In particular, should$ne: supersededretain both null and missing values?I'd also test a limit of 5 with superseded notes ranked ahead of at least five eligible notes, and check that the filtered call still returns five eligible results. That would catch filtering only after the candidate limit has been applied. The per-call historical override could then be checked against the same fixture.
I haven't run this against Basic Memory; these are suggested acceptance tests, not a bug report.
kamal0808 commentedon Nov 15, 2026
I hit this running a company's memory on Basic Memory. What worked for me was fixing it at write time: one fact per observation line, each with a
sincedate and a source. When a value changes, the old line stays, tagged#replacedwith a pointer to the new one, so each property has exactly one#currentline.On your top-5 point: this doesn't stop old lines from ranking. Replaced lines still come back in search (I search with
items: ["observation"]). What changes is that every hit says whether it's current, so the model answering can skip the stale ones. I haven't tested filtering on the tag itself.On LongMemEval's 78 knowledge-update questions (Haiku answering, Sonnet judging), the whole brain in the prompt scored 69/78 vs 63/78 for reading the full history. After a fix to the writing rules I re-ran 71 of those questions, answering only from BM's top 15 hybrid search hits (top 15, not your top 5): 64/71, against 65/71 with the whole brain in the prompt. Self-run, one answer model, one judge, so take it as a signal.
The writing rules are a free Claude skill: https://github.com/topc-social/brainy. @Flynn-Kawolsky, since you already have a retrieval eval, would you try it and tell me if stale facts stop reaching the answer? Even a "no change" would help.