feat: persistent per-user memory with pgvector #13

Closed

Jc-965 wants to merge 0 commits from persistent-user-memory into main

pull from: persistent-user-memory

merge into: ScottyLabs:main

ScottyLabs:main

ScottyLabs:repo-restructure

ScottyLabs:feat/email-integration

Conversation 1 Commits - Files changed - +4525 -478

Jc-965 commented 2026-08-06 04:57:26 +00:00 (Migrated from codeberg.org)

Copy link

Summary

Signed-in users now have durable memory across chats. The agent remembers facts a user shares or explicitly asks it to keep, and recalls them in later conversations.

• Facts are stored in PostgreSQL with pgvector (text-embedding-3-large, halfvec(3072) HNSW index) and recalled by semantic similarity each turn; without DATABASE_URL the store falls back to in-memory for dev and CI
• The model gets remember/forget tools, and a rate-limited background pass distills durable facts from normal conversation; raw chat turns are never stored
• GET/DELETE /memory/{user_id} endpoints power the Surface's memory manager: substring search, filter by learned vs remembered, per-item delete, clear all
• Memory is namespaced per user behind the shared secret; user ids are allowlisted so they can never form a cross-tenant LIKE pattern, recalled facts are wrapped as untrusted data, and memory tools are trusted by identity rather than name, so an MCP tool named remember stays untrusted
• /api/health reports the memory backend, readiness, and whether semantic search is active
• Moderate user messages and agent replies

### Summary Signed-in users now have durable memory across chats. The agent remembers facts a user shares or explicitly asks it to keep, and recalls them in later conversations. - Facts are stored in PostgreSQL with pgvector (`text-embedding-3-large`, `halfvec(3072)` HNSW index) and recalled by semantic similarity each turn; without `DATABASE_URL` the store falls back to in-memory for dev and CI - The model gets `remember`/`forget` tools, and a rate-limited background pass distills durable facts from normal conversation; raw chat turns are never stored - `GET/DELETE /memory/{user_id}` endpoints power the Surface's memory manager: substring search, filter by learned vs remembered, per-item delete, clear all - Memory is namespaced per user behind the shared secret; user ids are allowlisted so they can never form a cross-tenant LIKE pattern, recalled facts are wrapped as untrusted data, and memory tools are trusted by identity rather than name, so an MCP tool named `remember` stays untrusted - `/api/health` reports the memory backend, readiness, and whether semantic search is active - Moderate user messages and agent replies

Jc-965 added 1 commit 2026-08-11 01:19:11 +00:00

docs: improve comments for persistent memory 8074ca92bc

scottylabs-bot commented 2026-08-11 01:21:09 +00:00

Owner

Copy link

Kennel Deployments

Torn down at 2026-08-13 04:12:23 UTC.

<!-- kennel-deployment --> ### Kennel Deployments Torn down at 2026-08-13 04:12:23 UTC.

Jc-965 added 1 commit 2026-08-11 04:57:12 +00:00

feat: make memory forgetting more robust 7e390baac4

Jc-965 changed title from WIP - feat: persistent per-user memory with pgvector to feat: persistent per-user memory with pgvector 2026-08-11 04:57:36 +00:00

Jc-965 added 1 commit 2026-08-12 00:14:14 +00:00

feat: add chat title generation endpoint 13dba7a277

Jc-965 added 1 commit 2026-08-12 05:23:13 +00:00

feat: moderate user messages and agent replies 27dc1c6122

Jc-965 added 9 commits 2026-08-13 04:04:17 +00:00

feat: let the model decide the campus map via maps_show_map 30234c5e5b

feat: cut per-turn input tokens and add daily per-user token budget b23c0a8446

feat: enforce safety in code and further improve token optimization 5dd98ae0d1

fix: improve caps and remove global limit 48f25d5b3e

docs: improve comment flow b1afe89a66

fix: removed leaked reasoning delimiters 9a17e07d2b

fix: improve map tool-failure handling and map comments 36ba164c6d

Merge remote-tracking branch 'origin/cmumaps-improvements' into persistent-user-memory 9f2214c36e

Merge remote-tracking branch 'origin/optimization-token-limit' into persistent-user-memory 2ee505d85d

Jc-965 closed this pull request 2026-08-13 04:12:22 +00:00

Jc-965 referenced this pull request 2026-08-13 04:19:58 +00:00

feat: persistent memory, moderation, token budgets, and model-driven maps #16

Some checks failed

Hide checks

CI / build (pull_request) Failing after 16s

Details

CI / check-1 (pull_request) Failing after 18s

Details

CI / check (pull_request) Failing after 0s

Details

CI / test-agent (pull_request) Has been skipped

Details

kennel/build build succeeded

Details

kennel/deploy deployment healthy

Details

Pull request closed

This pull request cannot be reopened because the branch was deleted.

Sign in to join this conversation.

Reviewers

No reviewers

Labels

Clear labels

  

needs/docs

Must have documentation written

  

needs/research

Discuss the ideal implementation with the tech lead

  

needs/upstream

Needs a contribution to an upstream project

  

type/bug

Something not working as intended

  

type/external

Lives outside of code

  

type/feature

New functionality

  

type/refactor

Improving existing code

  

urgency

high

This is currently causing issues

  

urgency

immediate

This is making a project or feature unusable

  

urgency

low

This can happen whenever

  

urgency

medium

This is expected to cause issues soon

No labels

needs/docs

needs/research

needs/upstream

type/bug

type/external

type/feature

type/refactor

urgency

high

urgency

immediate

urgency

low

urgency

medium

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

3 participants

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference

ScottyLabs/cmugpt-agent!13

No description provided.