Skip to content

memory-recall-vs-static.md

Latest commit

 

History

History
41 lines (29 loc) · 1.75 KB

memory-recall-vs-static.md

File metadata and controls

41 lines (29 loc) · 1.75 KB

Memory Recall vs Static CLAUDE.md

Date: 2026-04-08

This is the first checked-in micro-benchmark answering a practical product question:

If you do not want a full cognitive runtime, how far can a static CLAUDE.md take you?

Baselines

Baseline Description
nexo_full_stack Shared brain, tools, learnings, workflow runtime, guardrails
static_claude_md Large manually maintained CLAUDE.md, no MCP memory tools
no_memory Fresh session, no persistent memory

Results

Scenario NEXO full stack Static CLAUDE.md No memory
Decision rationale recall Pass Partial Fail
User preference recall Pass Partial Fail
Repeat-error avoidance Pass Partial Fail
Interrupted-task resume Pass Partial Fail
Related-context stitching Pass Fail Fail

Notes

  • NEXO wins hardest on the scenarios that require tool-backed retrieval plus durable execution state.
  • A static CLAUDE.md can carry high-level style and some preferences, but it degrades quickly on:
    • rejected-alternative recall
    • file-scoped prevention rules
    • interrupted work that needs explicit checkpoints
    • context that lives across reminders, followups, learnings, and prior operational work
  • no_memory fails every scenario because the prompts intentionally omit the answer.

Why this benchmark matters

This benchmark is deliberately smaller than LoCoMo. It exists because users often compare NEXO not against other research systems, but against the simpler habit of stuffing more instructions into one bootstrap file. The answer from this first run is:

CLAUDE.md is useful context. It is not a substitute for a shared brain.