Memory Management

Memories hold confirmed knowledge that remains useful beyond one conversation or one task. Use TokST Cloud MCP to retrieve and write memories; use Dashboard → Memories to browse, edit, review, archive, and trace each record.

What Belongs in Memory

Type	Store
`fact`	Verified project, user, infrastructure, or operational information
`decision`	An approved choice, its rationale, and constraints
`preference`	Durable personal, team, or project conventions
`architecture`	System boundaries, components, interfaces, and operational design
`note`	Reusable context outside the typed categories
`task`	Historical compatibility records; assigned work belongs in a formal Task

Keep execution results in a Session or Task. Extract one or more Memories only when the result contains knowledge that another person or Agent can reuse later.

Agent Workflow

Call tokst_context for a compact scoped view, or call tokst_search for a specific question.
Compare the result with current evidence before proposing a change.
Call tokst_remember after a fact or decision is confirmed.
Include a title, suitable type, focused tags, and structured Markdown when it improves review.
Update, archive, restore, or supersede the existing record when knowledge changes.

The MCP connection provides the authenticated user and trusted Agent identity. Select the narrowest Workspace and Atlas that match the work.

Write High-Quality Records

Use short plain text for a simple fact. Use Markdown when it preserves material structure:

Type	Useful sections
Fact	Conclusion, source, scope
Decision	Decision, context, rationale, impact
Preference	Preference, when it applies, exceptions
Architecture	Goal, components, data flow, constraints
Note	Summary, details, next action

Preserve important values, tables, configuration changes, file paths, confirmed risks, and concrete next steps. Keep the working language used by the user and source material.

Cloud MCP accepts up to 512 KiB of UTF-8 Markdown per memory. Successful writes return the saved byte count. When a source is larger, attach the original file and store a complete, readable summary with its source link.

Data That Stays Outside Memory

Do not save secrets, passwords, API keys, private keys, cookies, or access tokens. Keep raw reasoning, terminal logs, raw tool JSON, debug traces, quota errors, network retries, and duplicated output in the active work record or diagnostics.

Preserve user-authorized business and personal information needed by the record. Ordinary writes skip automatic sensitive-content classification. Optional encrypted sensitive storage is disabled by default; the workspace owner can enable it in Dashboard Settings for explicitly selected manual records.

Search and Context

TokST uses keyword-first retrieval with semantic fallback. Direct text matches return immediately. Semantic retrieval is available when an embedding provider is configured and keyword results do not answer the request.

Use tokst_search for a focused question, type filter, tag filter, or historical decision. Use tokst_context when an Agent needs a compact grouped view of relevant facts, decisions, preferences, architecture, and notes before work.

Lifecycle and Evidence

State	Meaning	Retrieval	Recovery
Active	Current and usable	Included by default	Available
Archived	Kept out of normal retrieval	Excluded by default	Restore available
Deleted	Permanently removed	Unavailable	Unavailable

Use evidence, confidence, validity date, and review status for records that need a clear trust trail. A corrected decision should preserve its predecessor through an update or supersession link, so later Agents can understand the change.

Files and Import

Attach supporting files in the Dashboard or through the relevant Cloud MCP attachment tools. Use imports for established source material, then review the resulting records before treating them as durable knowledge.

Cloud MCP is the active memory interface. The CLI operates connected Agent endpoints, sessions, tasks, and diagnostics.