Skip to content

benchmark.md

Latest commit

 

History

History
176 lines (123 loc) · 6.59 KB

benchmark.md

File metadata and controls

176 lines (123 loc) · 6.59 KB

git-memory Benchmark & Comparison

Benchmarked April 13, 2026 on Windows / Git Bash

Validation of git-memory against simple-memory MCP — the most widely used MCP-based memory system. Tests run against a real corpus of 800+ memories migrated from simple-memory.

Migration

Metric Result
Memories migrated 800+
Date range 8 months
Unique tags 1,500+
Method git fast-import (single process)
Time 0.3s
Data loss None — all content, tags, and timestamps preserved

The initial approach (one git commit subprocess per memory) ran at ~3/s. Switching to git fast-import achieved 2,600+/s.

Search Quality

Identical queries run against both systems with comparable result limits.

Query simple-memory (limit:50) git-memory OR git-memory AND
cosmosdb 29 22 —
dri incident 30+ noisy (broad matches) 7 precise
certificate tls 20+ 9 broad 9 precise
memory mcp 20+ 14 —

Observations

  • simple-memory returns relevance-ranked results. Defaults to 10; must pass limit:50 for full recall.
  • git-memory OR returns all matches unranked. Broad terms pull in noise.
  • git-memory AND (+word1 +word2) is the precision tool. Results are clean and specific.
  • Recall quality is equivalent when both systems are tuned (AND in git-memory, limit:50 in simple-memory).

Performance at Scale (800+ memories)

Operation Before fixes After fixes Speedup
AND search 72.6s 0.7s 100x
add with dedup check hung indefinitely 0.9s ∞
OR search (single keyword) 0.4s 0.4s —
OR search (rare keyword, 56 results) 0.6s 0.6s —
Search (no results) 0.4s 0.4s —

Root causes fixed

  1. AND search spawned one git log -1 subprocess per candidate commit. Fix: git log --all-match --grep=X --grep=Y — one process, sub-second.
  2. Dedup check iterated all commits × all words in a nested bash loop. Fix: git log --grep per word + awk frequency counting.
  3. OR dedup used O(n²) bash array scan. Fix: awk '!seen[$1]++' pipe.

Agent Round-Trip

Store a memory, then search and retrieve it — the core agent workflow.

Store

simple-memory MCP git-memory
Mechanism 1 MCP tool call 1 terminal command
Time instant (network-bound) 0.8s (local, includes dedup)

Search

simple-memory MCP git-memory
Mechanism GraphQL query git-mem search --json
Time instant 0.17s (text) / 0.28s (JSON)
Output {"data":{"memories":[...]}} [{"hash":"...","date":"...","subject":"..."}]

Full content fetch

simple-memory MCP git-memory
Mechanism memory(hash:"...") git-mem show --json <hash>
Time instant 0.08s (text) / 0.27s (JSON)
Output JSON with content, tags, createdAt JSON with hash, date, subject, body

Token efficiency

simple-memory MCP git-memory --json
Search (7 results) ~820 chars ~1,100 chars
Show (full body) ~1,450 chars (title duplicated) ~1,100 chars (no duplication)

Cold Start

Time from zero (no install) to first memory stored and searchable.

simple-memory MCP git-memory
Prerequisites Node.js, npm git, bash
Install steps npm install + MCP config + VS Code restart curl one file + chmod
VS Code restart Required Not required
Time to first memory Minutes 0.6s
Works on SAW/devbox ⚠️ Needs Node.js ✅ git is pre-installed
Works in CI Needs MCP runtime ✅ Just bash
Works offline ❌ (MCP server dependency) ✅
Works on Termux (mobile) ⚠️ ✅

Unit Tests

54 tests across 4 suites, all passing.

Suite Tests Coverage
test-basic.sh 25 init, add, tags, stats, export, show, recent, help, errors
test-dedup.sh 6 duplicate detection, false positives, short messages, --yes mode
test-search.sh 12 OR, AND, case insensitivity, tag search, dedup, edge cases
test-sync.sh 11 remote setup, pull/push, divergent history, data loss check

Auto-Capture Quality Audit

Audited all memories tagged [auto] in a production simple-memory MCP store (10 memories, one week).

Category Count %
Genuinely valuable 2 20%
Context dump (stale quickly) 2 20%
Trivial/noise 5 50%
Duplicate 1 10%

Pattern

The 2 valuable auto-saves both corrected a wrong mental model — the kind of thing you'd want to remember. The 5 noise entries were ephemeral session activity the agent happened to observe. The heuristic doesn't distinguish "user discovered something important" from ambient activity.

Impact on the comparison

Silent auto-capture (often cited as an MCP advantage) is actually a liability at this signal-to-noise ratio. Terminal-visible saves (git-memory) let you notice and reject noise in real time. SKILL.md auto-capture rules are tightened to require a higher bar for [auto] tagged saves.

Where git-memory Wins

  • Zero dependencies — works anywhere git exists
  • Offline-first — no server, no network needed
  • Cold start — 0.6s from nothing to first memory
  • Transparency — every memory is a git commit, inspectable with any git tool
  • Portability — SAW, devbox, Termux, CI, bare Linux
  • No compile/rebuild cycle — edit SKILL.md or the bash script and you're done. MCP needs TypeScript rebuild + server restart.
  • Visible auto-capture — terminal commands let you catch noise. Silent MCP saves accumulate junk unnoticed.
  • Dedup built-in — warns before storing near-duplicates. MCP has none.
  • Cross-agent compatibility — SKILL.md works in VS Code Copilot, Claude Code, Cursor, any agent that reads files. MCP only works in MCP-compatible clients.
  • No maintenance — no server process to monitor, restart, or debug.

Where simple-memory MCP Wins

  • Structured API — native tool calls, no terminal parsing
  • Relevance ranking — results scored by relevance, not just grep matches

Where They're Equal

  • Recall quality (when tuned: AND search vs limit:50)
  • Token efficiency (~same chars per result)
  • Full content retrieval (identical data)
  • JSON output (git-memory --json matches MCP structured responses)

Not yet tested: failure recovery, parallel write safety, cross-platform (macOS/Linux/WSL2).