Memory store and recall

[owainlewis]

Summary

Add a memory capability with a NoMemory default, session summaries, and recall support for durable project knowledge.

Motivation

Sessions persist transcripts, but Neo does not yet distill reusable knowledge across sessions. Memory should preserve decisions, preferences, and project facts without bloating every prompt.

Scope

• Introduce internal/memory with a small store interface and NoMemory default.
• Add a file-backed implementation for session summaries and indexed recall.
• Add a bounded memory preamble for high-value recent context.
• Add a recall tool only after the storage shape is stable.
• Keep memory distinct from session transcript persistence.

Acceptance criteria

• Memory can be disabled with no behavior change.
• Session summaries are written in a human-readable local format.
• Recall returns bounded, relevant entries.
• Tests cover no-op memory, file-backed save/recall, and prompt preamble bounds.

Agent Assessment

Risk: high
Type: feature
Agent-ready: no

Reason:
Memory affects persistence, prompt context, privacy/security, and retrieval quality. Bad memory is worse than no memory, so this needs deliberate human-led design.

Suggested plan:

1. Define the memory store interface and human-readable storage format.
2. Decide what is summarized, recalled, and injected into prompts.
3. Verify no-op behavior, bounded recall, and prompt preamble limits.

Human needed:
Decide the first memory scope: session summaries only, project facts, or recall tool support.

[owainlewis]

Human decision for a first implementation slice: memory V1 should be manual and project-local. It should add features.memory, /memory <text>, project-root memory.md, and prompt loading when enabled, with no autonomous memory-writing tool yet. Agent-ready child issue: #117.