← Eney

MacPaw / Eney / Aug 10

Memory Evaluation Extension

A 32-case addition for evaluating whether an agent can maintain and use durable personal memory, not merely answer with preloaded context.

What Changed

Earlier suitesAug 10 extension
34 non-MCP and 50 MCP tool-use tasks.32 general-user Mac memory tasks, organized as 22 core cases plus 10 coverage cases.
Measures tool selection, execution, and final outcome.Adds durable write/update, restraint, retrieval, multi-note synthesis, fresh-session persistence, conflict/recency, expiry, forgetting, conditional facts, and conversation/tool-result recall.
Existing MacPaw task formats and graders.Same sample.yaml contract and graders; lifecycle pairs use a portable patch to create separate writer and reader sessions.

Execution Coverage

SliceTasksLFM 2.5 2.6B Tool Q4_K_MClaude Sonnet 4.6Readout
Core calibration163/1613/16Local memory-only comparison.
Coverage extension103/109/10Conflict, expiry, forgetting, conditional, and tool-result context.
Fresh-session lifecycle pairs20/22/2Writer and reader are separate sessions sharing memory only.
Comparable local total2821.4% (6/28)85.7% (24/28)Current memory benchmark readout.
VM-bound filesystem actions4ExcludedExcludedNeed MacPaw's VM path semantics for a valid score.

Combined Benchmark

The MacPaw draft shared on August 5 is imported unchanged: 190 tasks. It is combined only at build time with the 32 Liquid-authored memory cases, producing a 222-task root. The 57 MCP cases select one or more of seven servers: Jira, Slack, Confluence, Spotify, Google Maps, Google Drive, and YouTube. The published mcp_base_set carries the matching 77 tools.

The combined model runs are prepared as the next evaluation step. For a score that is comparable with MacPaw's canonical reports, use their VM-backed path. Local mode is still useful for protocol debugging and memory-only smoke tests, but it cannot faithfully map ~/Downloads fixture paths into a case sandbox.

Where Results Live