Agent Memory Discipline

Agent Memory Discipline supplies standing rules for using an existing memory tool. Its starting point is a practical failure: connecting a tool does not ensure that an assistant recalls decisions or saves corrections. The skill makes those moments explicit so project conventions and preferences can carry between sessions. It defines a sequence of recall, save, entry formatting, supersession, contradiction handling, and evidence assessment.

Recall comes before returning to a project, choosing libraries or patterns, writing tests or documentation where conventions matter, and answering questions about previous agreements. Searches use the user's words together with the project or repository name. An unhelpful search permits one broader query; repeated empty searches do not become a loop. Self-contained factual questions and arithmetic can skip recall. Large recall results are filtered to the current project and task.

Saving follows durable decisions, corrections, failed approaches with their reasons, continuing preferences, and environment facts learned through experience. Each entry holds one fact, making later supersession manageable. Entries record what changed, why, when it became true, and the source of the evidence. Record the scope of the decision as well as its evidence source, using separate backend fields when available. Evidence and policy entries can be explicitly labeled. User wording is preferred where paraphrasing could change meaning.

Old decisions remain available after they stop applying. The rule is to close the earlier entry with an end date and a pointer to its replacement rather than erase its historical explanation. Conflicting entries are shown with their dates and sources so a user can resolve them. Evidence from one run remains distinct from policy; human confirmation, a merged decision record, or repeated success is required before an observation becomes a rule.

Use this skill when a memory tool is connected but used inconsistently, when an assistant loses preferences between sessions, or when persistent project memory needs operating rules. Requests to remember a decision or recall an agreement are also direct triggers. Avoid saving transient state, readable file contents, temporary instructions, secrets, or personal data. This discipline does not replace environment validation, testing, or expert review.

The input is the current task together with memories returned by the chosen backend. Before work starts, the output is relevant recalled context identified by date and source. After a decision, correction, or failure, the output is a new single-fact entry. Superseded entries retain their text and receive closure information. This example illustrates the expected entry shape:

Project uses pnpm, not npm. Stated by the user on 2026-08-11 after a lockfile conflict. Applies to all packages in this repo.

If the tool fails, the assistant proceeds without memory, reports the failure once, and saves the pending decision when access returns. Rejected saves are reported rather than repeatedly rephrased.

The rules support Markdown files, local MCP memory servers, and hosted memory services over MCP. Authentication belongs to the selected backend. Removal requests are handled through backend-supported closure or deletion, with confirmation of what was removed. The skill installs no backend, handles no credentials, and performs memory operations through the tool already connected. Hosted services place entries outside the local machine, a consideration when selecting storage. License: CC0-1.0.