feat(memory): add holographic fact store and MCP tools - #11
Merged
Conversation
|
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 487f76fa08
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
added 5 commits
June 8, 2026 01:13
…phic-memory-core # Conflicts: # src/hooks.rs
added 8 commits
June 8, 2026 01:47
Eliminate the per-id N+1 round-trips in the holographic search/probe/reason paths by adding batched MemoryStore::get_facts and fact_vectors (chunked at 256 to stay under SQLite's 999-parameter limit). search now hydrates missing candidates and preloads candidate vectors in single batched queries, and holographic scoring becomes a pure helper fed those preloaded vectors; results_for_fact_ids does one batched fetch while iterating the original id order so probe/reason ordering is preserved. Worst case drops from ~1,100 round-trips to ~6. Also bind the user-derived strings in reason and entity_candidates as anonymous ? placeholders instead of interpolating string literals (escape_like still governs LIKE wildcards, now bound rather than inlined), and factor the repeated BEGIN IMMEDIATE/COMMIT/ROLLBACK boilerplate in add_fact, update_fact, remove_fact, and record_feedback_event into a with_immediate_tx combinator. Behavior is unchanged; all memory tests pass.
Cursor transcript ingestion previously re-read the whole transcript and re-upserted every message on every beforeSubmitPrompt, which is O(n^2) over a session under the 5s hook budget. Make ingestion incremental: resume from the byte offset tracked in the project session DB's parse_offsets table (mirroring the Claude accounting parser), seek + stream only newly-appended lines, and upsert just those. Handle truncation/rewrite (restart at offset 0) and a partial final line at EOF (defer until newline-terminated). Session start time and title are preserved across appends; ended_at advances to the latest message. Move the bulk catch-up off the hot path: register a Cursor stop hook (end-of-turn, generous timeout) as the primary ingest point and also ingest on sessionStart for resumed sessions. beforeSubmitPrompt now does at most a small tail read, bounded by a byte cap and wrapped in tokio::time::timeout so it fails open and can never exceed the prompt budget; the ingest runs inline (not a detached task) since CLI hook processes exit immediately. Upserts stay idempotent via INSERT ... ON CONFLICT, so concurrent short-lived hooks converge to the same rows. Behavioral change: re-ingesting an unchanged transcript is now a no-op, and the message ordinal/fallback id derives from the line's byte offset rather than a dense line index. Tests updated/added for the no-op, incremental append, and partial-line cases, plus the new stop-hook registration.
Introduce a shared TranscriptSource driver that persists ByteOffset, ContentHash, and RowCursor progress through the existing parse_offsets table, then refactor Cursor ingestion onto it without changing its hook behavior. The driver centralizes idempotent session/message upserts, source offsets, partial-line handling, full-rewrite content hashes, and row-cursor support. Add serve-side catch-up ingestion for hookless, path-discoverable adapters and implement the feasible sources verified in this pass: Claude Code JSONL, Codex rollout JSONL, Mistral Vibe messages.jsonl, and the Cline/Roo/Kilo full-rewrite task-history stores. Gemini transcript ingestion is intentionally omitted per user request; Hermes remains excluded. Add focused fixture tests for every integrated adapter plus framework tests for ByteOffset, ContentHash, and RowCursor cursor behavior.
Ensure Cline-family transcript fixtures are written under the same OS-specific VS Code data root that the provider scans, so macOS CI exercises the intended ingest path.
ScriptedAlchemy
changed the base branch from
split/holographic-cursor-local-base
to
master
June 8, 2026 02:10
Align message-search providers with ingested session sources, bound transcript storage and catch-up reads, and ensure memory transactions roll back when COMMIT fails.
This was referenced Jul 3, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Test plan
Stacked on #10.