dsh-mnemos

Once a DSH session ends, the model has no recollection of the conversation—repeating environment constraints and correcting the same facts wastes tokens and patience. dsh-mnemos adds a governed cross-session memory layer: every write, whether from model tools, /memory commands, or third-party plugins, must pass a single auditable approval gate. Sensitive, duplicate, or over-budget writes are rejected outright; third-party memories are flagged as unverified and capped at injection time, keeping the store safe from poisoning.

Retrieval uses an FTS5 multi-level word ladder fused with bigram similarity via RRF, scoring 87.2% hit@1 on LongMemEval-S. During a session only a one-line-per-memory index is injected (byte-stable, KV-cache friendly), and the model drills into details on demand. Memories are not static: whole buffered conversations can be batch-distilled into candidates, facts update in place (old values remain recoverable via Git), genuine conflicts generate replacement proposals for human review, and active memories can be formalized into portable DSH SKILL files. Heat-based eviction archives stale entries without hard deletion, while pinned memories stay active indefinitely.

For developers, ctx.mnemosBus exposes an open memory bus: any DSH plugin mounts it and calls recall, record, or subscribe. Third-party writes must declare an identity and always enter the human approval queue; a runtime blacklist and per-plugin revocation provide kill switches—sharing is permissioned, not assumed. Data lives in two tracks: SQLite (WAL + FTS5) for fast retrieval, and every memory is also a Markdown file in a Git repo, giving diff, rollback, backup, and cross-device push/pull sync for free. It suits developers and power users who want persistent, auditable, governed memory in DSH with cross-session reuse of experience and constraints.