recall-context returns excessive data causing token limit warnings

[mdlopresti]

Problem

The recall-context tool retrieves all memories at once, returning massive amounts of data (196K+ characters) that trigger token limit warnings in Claude Code and make the tool impractical for agents with moderate memory usage.

Evidence

First reported: December 17, 2025
Re-confirmed: December 26, 2025 (v0.4.0)

Actual Test Results (Dec 26, 2025)

Error: result (196,595 characters) exceeds maximum allowed tokens. 
Output has been saved to file.

With only:

• 16 private longterm memories
• 6 team memories
• 0 personal memories

Total: 22 memories → 196,595 characters

Impact

Severity: MEDIUM 🟡

• Triggers "large context" warnings in Claude Code
• Makes recall-context impractical for regular use
• Agents hit context limits faster
• Inefficient use of context window
• Forces workarounds (reading from saved files, manual filtering)

Current Behavior

recall-context returns ALL memories from requested scopes in a single response:

• All private memories
• All personal memories
• All team memories
• All public memories
• Complete content for each memory
• All metadata (tags, priority, timestamps, etc.)

This is a dump everything approach that doesn't scale.

Expected/Desired Behavior

Phase 1 (Quick Fix):

• Add limit parameter (default: 20) ✅ ALREADY EXISTS
• Document that users should use limit parameter
• Maybe reduce default from current behavior

Phase 2 (Better Architecture):

• Return index/summary first (memory IDs, titles, tags, truncated content)
• Allow retrieval of individual memories by ID on demand
• Support pagination for large memory sets

Root Cause

Pattern v0.4.0 enhanced recall-context with:

• Tag filtering
• Priority filtering
• Date range filtering
• Content search

But all filtering is client-side - still retrieves ALL memories first, then filters.

With moderate usage (20-30 memories), this becomes unusable.

Proposed Solutions

Option A: Limit by Default (Quick Fix)

1. Change default limit from unlimited to 50
2. Document limit parameter prominently
3. Add warning if limit reached: "Showing X of Y total memories"

Option B: Two-Phase Retrieval (Better UX)

1. recall-context returns summary/index with memory IDs, previews, tags
2. New tool: get-memory for full retrieval by ID

Option C: Smart Summary (AI-Friendly)

1. Generate 4KB summary at retrieval time
2. Group by category, prioritize by recency/priority
3. Include memory IDs for drill-down

Workaround

Currently, agents must:

1. Read saved file when recall-context exceeds limits
2. Manually parse/filter JSON
3. Use specific filters to reduce result set

Environment

• Pattern version: v0.4.0
• 22 total memories across private + team scopes
• Result: 196,595 characters (exceeded token limit)

Related

This is mentioned in Pattern memory as known issue (Dec 17), but was never formally filed as GitHub issue.