Feature: Allow disabling automatic memory recall (FTS5) per-message

[weissfl]

Problem

enrichMessageWithRuntime() runs FTS5 + BM25 recall on every incoming message, with hardcoded parameters:

• DEFAULT_RECALL_LIMIT = 5
• MAX_CONTEXT_BYTES = 4000
• SCOPED_RECALL_CANDIDATE_LIMIT = 64
• GLOBAL_RECALL_CANDIDATE_LIMIT = 64

There is no way to disable this behavior via config. auto_save: false prevents automatic memory writes, but recall still fires on every message — even when the memories table is empty, causing unnecessary SQLite queries.

Why this matters

1. Unnecessary overhead — FTS5 search runs even when memory is empty or the user does not need context injection. For high-throughput or professional use, this is wasted cycles.

2. Noisy context — Auto-recall can inject irrelevant memories into the prompt, degrading model quality. The user has no way to say "I will manage context myself."

3. Inconsistent with auto_save: false — If I can disable auto-writes, I should be able to disable auto-reads. Currently it is only half-controllable.

4. Hardcoded limits — recall_limit, max_context_bytes, and candidate limits are compile-time constants. Users with large memory stores or specific context budgets cannot tune them.

Proposal

Add config options under memory:

{
  "memory": {
    "auto_recall": true,
    "recall_limit": 5,
    "max_context_bytes": 4000
  }
}

• auto_recall: false — skip enrichMessageWithRuntime() entirely, no FTS5/LIKE queries
• recall_limit — configurable max results (currently hardcoded to 5)
• max_context_bytes — configurable context budget (currently hardcoded to 4000)

Manual memory_recall tool calls should still work regardless of auto_recall setting.

[weissfl]

Live reproduction: conversation degradation from auto-recall

The issue isn't just about performance — auto-recall actively degrades conversation quality.

On every message the runtime injected ~6 memory entries, none relevant to the current context. The model is forced to process them — result: confusion, loss of conversational coherence, wasted context window.

FTS5/BM25 returns top-N by keyword overlap, but keyword overlap ≠ conversational relevance. auto_recall: false isn't an optimization — it's needed to make the agent usable.