dsh-adaptive-context

Anyone who has worked with an agent has probably hit this frustration: you explicitly told it last week to keep config and code separate, yet this week it mixes them again; you worry about letting it remember, fearing an unreliable web snippet becomes a preference that taints every session; and when it remembers something wrong you cannot tell who recorded it, when, or how to fix it. The ACP (Adaptive Context Plane) plugin tackles exactly this problem by making agent memory behave like a ledger: every entry carries a source, an authority level, and an effective time window; writes pass through a governance gate, injection has a token budget, and behavioural changes require approval.

Memory is modelled across six explicit domains (user facts, preferences, work, expression style, experience, external facts), all grounded in an append-only SQLite evidence ledger where replaying the same event produces a single record. Context injection ranks candidates through both lexical and semantic channels, respects a per-turn token quota, and automatically deduplicates self-echoes. Background consolidation runs without blocking the conversation, uses an LLM to distil observations, retries on failure so no batch is lost, and falls back to rule-based extraction only when the LLM is unavailable. Expression-style promotion requires either human approval or a policy guardrail, backed by a five-state machine with full rollback. Every critical operation is audited, and a one-click JSONL export and import make migrating to a new machine painless.

It is built for developers who collaborate with an agent long-term and want memory that is controllable, auditable, and explainable, especially those who find the dump-everything-into-the-prompt approach too crude and are looking for a governance layer. If you are already using DeepSeek Harness, installing this plugin gives you a long-term memory system with gates, budgets, and audit trails out of the box, without building it yourself.