Problem

During long sessions, context compression removes older messages. Once compressed, those messages are:

1. Gone from active context (can't be referenced)
2. Not in MEMORY.md (requires explicit extraction)
3. Not searchable via session_search (only indexes past sessions)

This creates a "forgetfulness" gap mid-conversation. The agent cannot recall what was discussed earlier in the same session.

Example:

User discusses topic X at 09:30. By 15:30, context compression has removed those messages. If the user asks "what did we say about X earlier?", the agent cannot answer — the context is gone, and session_search won't help because the session hasn't ended yet.

Proposed Solution

A current-session buffer — a rolling archive of messages that were compressed out:

Features:

1. Searchable mid-session — tool that queries the buffer:
2. Lighter than MEMORY.md — not facts, just raw message history
3. Auto-cleanup — buffer is discarded when session ends (already in session_search)
4. Token-efficient — stored as text, not in active context

Implementation options:

Related

• Feature: Structured Memory System — Typed Nodes, Graph Edges, and Hybrid Search #346 — Structured Memory (extracts facts during compression, but requires LLM extraction)
• Feature: CLI Inspectable Tool Timeline — Expandable Tool Calls, Results, and Historical Event Inspection #1092 — CLI Inspectable Tool Timeline (preserves tool results, not messages)
• [Bug]: GLM gateway sessions can undercount request size, overflow late, and persist guessed fallback context limits #2599 — GLM context overflow handling (prevents overflow, doesn't preserve content)
• Feature: Two-Phase Context Management — Prune Tool Outputs Before Full Compaction (inspired by Kilocode) #513 — Two-Phase Context Management (better compaction ordering)

Alternative: Extend session_search

Current only indexes completed sessions. Could extend it to:

1. Index current session in real-time
2. Include "current session" results with a flag

But this changes session_search semantics and may not want all transient messages indexed permanently.

Use case: I lost context from 3 hours ago in this conversation. User asked "what did we discuss about Tavily?" and I couldn't recall because compression had already removed those messages from my context window.