[GH-ISSUE #38] Improve Memory System: Extend automatic vector search process for semantic memory (to complement episodic memory) #4

New issue

Open

opened 2026-10-26 22:06:43 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 22:06:43 +00:00

Owner

Originally created by @christian-byrne on GitHub (Jul 22, 2025).
Original GitHub issue: https://github.com/Comfy-Org/comfy-claude-prompt-library/issues/38

Current Memory System

We have:

Episodic memory: conversation sumamries
Semantic memory: knowledge bases which are iterated after every session where appropriate
Procedural memory: via commands, templates, hooks

At the start of each session, the user needs to prime the session with the memory relevant to the task.

Currently, the user is expected to start via knowledge exploration, which triggers agent to

(episodic) AUTOMATIC: do semantic search on conversation summaries automatically based on task keywords -> read up to 3 summaries with relevence score thresholds
(semantic) MANUAL: search the relevant knowledge folder -> create a topic index -> present users with the knowledge subtopics and ask which to hone in on. If user specifies sub topic(s), read the files.
(procedural) MANUAL: the user then uses the commands, hooks, and templates appropriaate to the given task

Proposed Improvement

Number 2 should happen automatically.

Currently we have this process: extracting topics from files in a folder -> presenting to user -> user select.

That process introduces a lot of error and (e.g., user forgets topics exist so skips past process, topic extraction is not done well because too many files and subfolders, etc.). Instead, the knowledge search can happen using same vector search and indexing system as the memories (via https://github.com/christian-byrne/claude-code-vector-memory).

Originally created by @christian-byrne on GitHub (Jul 22, 2025). Original GitHub issue: https://github.com/Comfy-Org/comfy-claude-prompt-library/issues/38 ## Current Memory System We have: - Episodic memory: conversation sumamries - Semantic memory: knowledge bases which are iterated after every session where appropriate - Procedural memory: via commands, templates, hooks At the start of each session, the user needs to prime the session with the memory relevant to the task. Currently, the user is expected to start via knowledge exploration, which triggers agent to 1. (episodic) AUTOMATIC: do semantic search on conversation summaries automatically based on task keywords -> read up to 3 summaries with relevence score thresholds 2. (semantic) MANUAL: search the relevant knowledge folder -> create a topic index -> present users with the knowledge subtopics and ask which to hone in on. If user specifies sub topic(s), read the files. 3. (procedural) MANUAL: the user then uses the commands, hooks, and templates appropriaate to the given task ## Proposed Improvement Number **2** should happen automatically. Currently we have this process: extracting topics from files in a folder -> presenting to user -> user select. That process introduces a lot of error and (e.g., user forgets topics exist so skips past process, topic extraction is not done well because too many files and subfolders, etc.). Instead, the knowledge search can happen using same vector search and indexing system as the memories (via https://github.com/christian-byrne/claude-code-vector-memory).

bsakaguchi referenced this issue

2026-10-26 22:06:48 +00:00

[PR #4] [MERGED] [ci] Add GitHub Actions validation workflow #8