Feature request: first-class user-owned structured memory (typed graph + retrieval discipline), consistent across surfaces · Issue #75291 · anthropics/claude-code · GitHub
//voltron/issues_fragments/issue_layout" data-turbo-transient="true" />
Skip to content
Search or jump to...
Search code, repositories, users, issues, pull requests...
-->
Search
Clear
Search syntax tips
Provide feedback
--><br>We read every piece of feedback, and take your input very seriously.
Include my email address so I can be contacted
Cancel
Submit feedback
Saved searches
Use saved searches to filter your results more quickly
-->
Name
Query
To see all available qualifiers, see our documentation.
Cancel
Create saved search
Sign in
//voltron/issues_fragments/issue_layout;ref_cta:Sign up;ref_loc:header logged out"}"<br>Sign up
Appearance settings
Resetting focus
You signed in with another tab or window. Reload to refresh your session.<br>You signed out in another tab or window. Reload to refresh your session.<br>You switched accounts on another tab or window. Reload to refresh your session.
Dismiss alert
{{ message }}
Uh oh!
There was an error while loading. Please reload this page.
anthropics
claude-code
Public
Notifications<br>You must be signed in to change notification settings
Fork<br>22k
Star<br>137k
Feature request: first-class user-owned structured memory (typed graph + retrieval discipline), consistent across surfaces #75291
New issue<br>Copy link
New issue<br>Copy link
Open
Open<br>Feature request: first-class user-owned structured memory (typed graph + retrieval discipline), consistent across surfaces#75291
Copy link
Labels<br>area:mcpenhancementNew feature or requestNew feature or requestmemory
Description
parrik<br>opened on Jul 7, 2026
Issue body actions
The one-paragraph version
The same user asking the same personal question gets confabulation from Claude on
one surface and grounded, checkable answers on another — and the difference is not
the model. It's whether the model has (a) a typed, user-owned memory substrate
and (b) an enforced retrieval discipline over it. Both exist today only if the
user hand-builds them (files + MCP + prompt rules) and only in Claude Code. They
should be product defaults, on every surface, including mobile.
What I built (as a user), and what it does
A plain-text knowledge graph the user owns and curates:
Typed nodes — observations / references / interpretations / practices /
forecasts — each with a stable ID, a one-line name, and a full statement.
Provenance on every node — how it was sourced (user-stated vs inferred),
evidence kind, a tentative flag, temporal validity, and dated corrections that
supersede older claims instead of silently overwriting them.
An entry-point node (current state) read first; everything else fetched by ID
on demand — the graph does not need to fit in context.
A local MCP search server whose semantic search deliberately returns names
and edges only, not statement bodies — finding a node and reading a node are
separate, explicit acts.
A cite-only-what-you-fetched rule — the assistant may not cite a node ID
unless that node's body is in the current turn's tool log; otherwise it must mark
the reference "name only, body unread."
Write gating in rings — low-stakes node types merge autonomously via a
staging queue; high-stakes types require an explicit human accept-click on a
review surface. The model proposes; the user curates.
Deterministic injection for hard rules (a hook that fires every turn) vs
relevance-recalled memory vs read-on-demand deep context — three channels with
different reliability guarantees for different classes of information.
What I observe in practice
With the graph (terminal + MCP): questions about my own history, patterns,
and decisions get answered from filed, dated, provenance-carrying records. When
the model overclaims, I correct it, and the correction gets filed with a date
and supersedes the old claim — the memory improves monotonically. In long,
emotionally loaded decision sessions, the assistant checks each framing against
the record instead of mirroring my mood, and it can tell me — accurately — when
today's framing contradicts what I said when I was clear-headed. That last
behavior is the single most valuable thing an AI assistant has ever done for me,
and it is entirely a function of the substrate.
Without the graph (phone app, same underlying models): the same questions
produce fluent, confident fabrication — plausible summaries of conversations
that drift from what was actually said, "memories" with no provenance, no
tentative-marking, no way to check. Summary-style memory sounds like knowing
me; typed memory with fetch-before-cite is knowing me, verifiably.
The asks, concretely
User-owned, schema'd memory as a product primitive. Not opaque
auto-summaries: typed records with IDs, provenance, tentative/confidence flags,
temporal validity, and dated supersession. Exportable,...