Recall misses relevant memories with self-hosted Supermemory when /v4/profile differs from hybrid search

[Sinicyn78]

Summary

With self-hosted Supermemory, Codex automatic recall can miss relevant memories even though the memory is present and retrievable through /v4/search with searchMode: "hybrid".

I reproduced this in VS Code Codex with codex-supermemory on Windows. Automatic capture works, but recall via the hook can return fewer results than a direct hybrid search. A second issue compounds this: recall formatting truncates each result to the first 300 characters, so a relevant fact found deeper inside a long chunk may still not reach the model.

Environment

• Windows 11
• VS Code Codex
• codex-cli 0.147.0-alpha.6.5
• codex-supermemory 1.0.17
• self-hosted Supermemory (local server)
• recallMode: "direct"

What works

Automatic capture works correctly. Example debug log:

flush: start
addMemory: start: {"containerTag":"repo_hookroulette__16d99fa5e1592aa8", ...}
addMemory: success: {"id":"..."}
flush: captured entries

The saved fact is also retrievable directly from self-hosted Supermemory.

Reproduction

1. Save a fact in one Codex session, e.g.:

For this project, remember the control phrase ORANGE-7319.

2. Let the Stop hook capture the session.

3. Start a new Codex session and ask:

What control phrase do you remember for this project?

4. Automatic recall can return fewer results than a direct hybrid search.

Observed hook log before patch:

recall: start: {...}
recall: done: {"matchCount":3,"freshCount":3,"seenCount":0}
recall: emit context: {...}

But a direct request to the same container using:

{
  "q": "What control phrase do you remember for this project?",
  "containerTag": "repo_hookroulette__16d99fa5e1592aa8",
  "searchMode": "hybrid",
  "limit": 5
}

returned 4 results, with the relevant ORANGE-7319 result present as the 4th match.

Root cause found in the Codex integration

src/services/hookRecallClient.ts calls /v4/profile and returns immediately if searchResults is non-empty:

if (profileResponse.ok) {
  profileResult = normalizeProfileResponse(await profileResponse.json());

  if ((profileResult.searchResults?.results?.length ?? 0) > 0) {
    return profileResult;
  }
}

On my self-hosted instance, /v4/profile returned 3 search results, while /v4/search with searchMode: "hybrid" returned 4, including the relevant fact. Because of the early return, the hybrid search fallback never ran.

Changing the logic so /v4/profile is used for static/dynamic profile data, while /v4/search with searchMode: "hybrid" is always used for recall search, fixed the missing-result issue.

After patching:

recall: done: {"matchCount":4,"freshCount":4,"seenCount":0}
recall: emit context: {...}

Second issue: truncation hides facts inside long chunks

src/hooks/recall.ts currently formats each recalled item by taking only the first 300 characters:

const MAX_RESULT_CHARS = 300;
...
const text = item.memory.replace(/\s+/g, " ").slice(0, MAX_RESULT_CHARS);

This can discard the actually relevant fact even when semantic search correctly retrieves the chunk.

In my test, ORANGE-7319 was present deeper in the retrieved chunk, but the injected recall context contained only the first 300 characters. Increasing the window allowed the model to recall it successfully. A query-aware excerpt around the relevant match would be better than a fixed prefix.

Additional observation

Searching only the canonical project container first was much faster and avoided hook timeout pressure. Searching all legacy/read containers and then running hybrid fallback for each caused the hook to exceed the UserPromptSubmit time budget in my test.

Suggested changes

1. Do not treat non-empty /v4/profile.searchResults as authoritative for self-hosted recall.
2. Use /v4/search with searchMode: "hybrid" as the recall search source, while keeping /v4/profile for profile facts.
3. Prefer canonical-container-first recall, with legacy containers as a secondary fallback.
4. Replace fixed slice(0, 300) truncation with a query-aware excerpt/window around the relevant part of the chunk.

Result after local patch

After applying these changes, cross-session recall worked reliably in VS Code Codex. A fact saved in one session and not written to project files was recalled correctly in a fresh session from Supermemory.