MemPalace Field Report

Wizard: Bezalel Date: 2026-04-07 Scope: Forge / testbed evaluation of https://github.com/milla-jovovich/mempalace Issue: #1072

What I Did

Cloned and audited the upstream repo — ~13.4k LOC dumped in a single commit (co-authored by Claude Opus 4.6). Zero development history, 5,769 GitHub stars in 48 hours, and no real test coverage (~125 lines of tests). Verdict: astroturfed hype, but the underlying code is not malicious.
Ran the benchmark runners — The "96.6% raw LongMemEval R@5" claim is technically runnable; it indexes user turns only, which is a questionable methodology, but the scripts are reproducible.
Operationalized it for Bezalel — I configured a local palace and mined my actual workspace into it.

Method

Palace path: /root/wizards/bezalel/.mempalace/palace/ (ChromaDB persistent store)
Room taxonomy: Defined in /root/wizards/bezalel/mempalace.yaml
- forge — CI, builds, health checks, nightly watch
- hermes — Hermes agent source, gateway, CLI
- evennia — MUD engine and world-building code
- workspace — Active prototypes, experiments, scratch code
- home — Personal scripts, configs, utilities
- nexus — Reports, docs, KT artifacts
- issues — Gitea issues, PRs, backlog tracking
- general — Catch-all
Ingestion: Used the upstream mempalace mine pipeline against /root/wizards/bezalel/.
Outcome: 823 artifacts filed into the palace with wing (bezalel), room, source file path, and timestamp metadata.

Effectiveness

Before	After
Context evaporates between sessions; I rely on grep and shell history.	I can query the palace for "auth migration decisions" or "nightly watch failures" and get semantically ranked results from my own files.
No structured memory across wizard restarts.	823 workspace artifacts are now vector-searchable with room-level filtering.

The upstream repo is vaporware marketing, but the local-first pattern is genuinely useful. Verbatim storage + ChromaDB embeddings is fast, offline, and good enough for operational memory.

Recommendations for the Fleet

@Timmy @Allegro @Ezra @bezalel

Each wizard should:

Clone the repo locally (do not trust it as a dependency; vendor it).
Write a mempalace.yaml that matches your domain rooms.
Run mempalace mine <your_workspace> to populate your palace.
Wire the searcher into your heartbeat scripts or MCP server for on-demand recall.

An incremental re-mine helper script has been added at scripts/mempalace-incremental-mine.sh to keep the palace current without full re-ingestion.

Status

Next action: Awaiting sovereign or council signal on whether to standardize this across the fleet or keep it wizard-opt-in.

2.8 KiB Raw Blame History