memory-bank（记忆库）

Agent Memory Bank

A persistent memory system for storing knowledge that survives across conversations — for any domain: code, business, personal, research, creative work, client management, and more.

Default location: ./memory/ (create if it doesn't exist)

Default structure:

• ./memory/episodic/
• ./memory/semantic/
• ./memory/procedural/

Core Philosophy

• Three memory types, one system: Store experiences as episodic, durable facts and evolving state as semantic, and repeatable workflows as procedural.
• Use the shape that matches future retrieval: Save things based on how you'll want to find them later — what happened, what is true, or how to do it.
• Prefer history over rewrites: When an event or decision matters in its own right, preserve it as a dated record rather than erasing what came before.
• But don't spam: Update living records when the goal is to maintain current state or current instructions.
• Never delete memory files — mark outdated ones with status: superseded instead.
• Write for resumption: Notes should be self-contained — a future session with zero context should still understand them.
• Bias toward usefulness: Save things that may matter later, especially details that are easy to forget but annoying to rediscover.
• Tags are first-class: Tag memories so they can be searched quickly across domains, projects, people, and topics with rg.

Quick Routing Cheat Sheet

Use this fast decision tree when deciding what to save:

• Ask: what will future-you want back?
• If the answer is what happened? -> save episodic
• If the answer is what is true now? -> save or update semantic
• If the answer is how do we do this? -> save or update procedural
• If the answer is more than one of those -> save it in more than one form

Mini examples:

• "We had a nasty auth incident today" -> episodic
• "The client prefers weekly async updates" -> semantic
• "Here is the release checklist" -> procedural
• "A failed deploy taught us a safer rollout order" -> episodic + procedural
• "A conversation revealed a lasting user preference" -> episodic + semantic

Memory Types

Episodic

Use for events, moments, decisions-in-context, meetings, incidents, milestones, discoveries, failed attempts, and progress snapshots.

This answers: what happened?

Examples:

• A debugging session that revealed the root cause
• A client call and the decisions made during it
• A travel day where something went wrong
• A negotiation update or milestone reached

Semantic

Use for durable knowledge, evolving facts, preferences, profiles, constraints, current project state, and anything that may be updated many times over.

This answers: what is currently true?

Examples:

• User communication preferences
• Current architecture constraints
• Client profile and standing preferences
• Active project status and known risks

Procedural

Use for repeatable methods, checklists, workflows, instructions, playbooks, routines, and standard ways of doing things.

This answers: how do we do this?

Examples:

• How to deploy safely
• Weekly reporting workflow
• Travel packing checklist
• How to onboard a new client

One interaction may produce more than one memory type. That's good.

Example:

• An incident happens -> save an episodic note
• It reveals a durable constraint -> update a semantic note
• It teaches a better workflow -> update a procedural note

Tags

Use tags generously but intentionally. Tags make cross-cutting retrieval easy without overcomplicating the folder structure.

Good tag categories:

• domain: code, business, personal, research, creative
• topic: auth, deploy, health, finance, planning
• entity: client-acme, project-helios, user, team
• type hints: decision, preference, incident, checklist, workflow

Prefer short, stable, lowercase tags with hyphens when needed.

Examples:

• tags: [code, auth, incident, project-helios]
• tags: [business, client-acme, preference, communication]
• tags: [personal, travel, checklist]

How to Route New Information

Ask these questions:

1. Is this mainly an event or time-bound moment?

- Save episodic

1. Is this a durable fact, preference, constraint, or current state?

- Save or update semantic

1. Is this a repeatable method or instruction?

- Save or update procedural

1. Does it fit more than one?

- Save it in more than one form when future retrieval would differ

Examples:

• "We debugged auth and found the cookie domain was wrong"

- episodic: the debugging session - semantic: auth depends on the correct parent-domain cookie setting

• "The client prefers weekly async updates"

- semantic: client preference

• "Here is our monthly reporting process"

- procedural: recurring workflow

• "Friday's failed deploy taught us to run migrations before workers"

- episodic: failed deploy - procedural: safer release workflow

Proactive Saving

Since this skill is loaded, memory is clearly valued here — lean into it. Save without being asked when you encounter:

• Something that took real effort to figure out (research, debugging, negotiation, comparison)
• A decision with non-obvious reasoning — why X over Y
• Information that would be painful to reconstruct if this conversation ended
• Dead ends and failed approaches — saves future sessions from repeating them
• In-progress work with clear next steps
• Anything the user seems to care about that isn't obvious from context alone

Applies to any domain — code, business, personal, creative, research, client work, anything.

File Naming

episodic:   YYYY_MM_DD_meaningful_name.md
semantic:   stable_subject_name.md
procedural: how_to_meaningful_name.md

Examples — notice these span many domains:

• 2025_03_09_auth_bug_root_cause.md
• 2025_03_10_q1_marketing_decisions.md
• user_preferences.md
• client_acme_profile.md
• project_helios_status.md
• how_to_rotate_api_keys.md
• weekly_client_reporting_workflow.md

Multiple episodic files per day are normal and encouraged. Semantic and procedural files should usually keep stable names so they can be updated over time.

Creating a Memory File

Use your native file tools to create and edit memory files.

Shared frontmatter fields:

---
summary: "One line — specific enough to know if you need to read this"
created: YYYY-MM-DD
updated: YYYY-MM-DD
memory_type: episodic | semantic | procedural
tags: [optional, tags]
---

Episodic template:

---
summary: "Debugging session that found the auth cookie domain mismatch"
created: YYYY-MM-DD
updated: YYYY-MM-DD
memory_type: episodic
tags: [code, auth, debugging, incident]
---

# Title

[Write the event clearly enough that a future session understands what happened and why it mattered]

Semantic template:

---
summary: "Current auth configuration constraints for production"
created: YYYY-MM-DD
updated: YYYY-MM-DD
memory_type: semantic
tags: [code, auth, configuration, project-helios]
---

# Title

[Write the current known state, preferences, constraints, or facts that should remain current over time]

Procedural template:

---
summary: "Safe production deploy workflow"
created: YYYY-MM-DD
updated: YYYY-MM-DD
memory_type: procedural
tags: [code, deploy, workflow, operations]
---

# Title

[Write the repeatable process clearly enough that someone can follow it later]

Some section ideas (use only what fits):

• Context — why this matters, background
• Key Decisions — what was decided and why
• Details / Findings — the actual content worth saving
• Current State — the up-to-date truth right now
• Procedure / Checklist — ordered steps or repeatable instructions
• Lessons Learned — what this teaches for future work, mistakes to avoid, or patterns worth reusing
• People / Contacts — who's involved
• Next Steps — what to do next
• Didn't Work — dead ends to avoid

There's no required structure. A memory for a client call looks different from a debugging session — that's fine.

Editing an Existing Memory

Use your native file editing tools to update memory files.

Update episodic when:

• You're still actively adding details to the same event or session
• You want to complete or correct the record of what happened

Update semantic when:

• The current state, preference, fact, or constraint has changed
• You learned something that refines a long-lived record
• You want one stable file to represent the latest understanding

Update procedural when:

• The workflow changed
• A checklist was improved
• You discovered a safer, faster, or clearer way to do the task

When a change deserves both history and current state, do both:

• create or update an episodic file for the event
• update the relevant semantic or procedural file for the lasting takeaway

Searching Memories

Quick orientation (run this first when memories feel relevant)

rg --files ./memory | sort

This shows the files that exist across all memory types — a fast orientation to what's here.

Search by memory type

rg "^memory_type: episodic$" ./memory/ --no-ignore
rg "^memory_type: semantic$" ./memory/ --no-ignore
rg "^memory_type: procedural$" ./memory/ --no-ignore

Read summaries across all files

rg "^summary:" ./memory/ --no-ignore

Search by keyword (full text)

rg "keyword" ./memory/ --no-ignore -i

Search summaries only

rg "^summary:.*keyword" ./memory/ --no-ignore -i

Search by tag

rg "^tags:.*keyword" ./memory/ --no-ignore -i

Search by tag combination

rg "^tags:.*auth.*project-helios|^tags:.*project-helios.*auth" ./memory/ --no-ignore -i

Search semantic memories for current state

rg "keyword" ./memory/semantic/ --no-ignore -i

Search procedural memories for how-to guidance

rg "keyword" ./memory/procedural/ --no-ignore -i

Search episodic memories for what happened

rg "keyword" ./memory/episodic/ --no-ignore -i

After finding relevant files, read them using your native file tools. The summary tells you if it's worth reading; reading gives you the actual context.

When to Check Memories

Default behavior: check memories when the task feels like it might connect to prior work — e.g. the user references a past decision, says "like we discussed", or you're about to re-research something familiar.

This default can be overridden by the environment. A system prompt may instruct you to always check memories, or only check when explicitly asked — follow those instructions. This skill describes the fallback when no such instruction exists.

When checking:

• orient quickly to what exists
• choose the memory type that matches the question
• search by tags, keywords, and summaries
• read the most relevant files for actual context

Use this retrieval pattern:

• episodic for "what happened?"
• semantic for "what do we know now?"
• procedural for "how do we do this?"

When to Save

No strict rules — use judgment. Good candidates:

• Something that took real effort to figure out
• A decision with non-obvious reasoning behind it
• Information you'd lose if this conversation ended now
• Anything the user explicitly wants remembered
• A preference, profile, or constraint likely to matter again
• A workflow or checklist you'll likely reuse
• A bug, config detail, or environment-specific gotcha that could bite again later

Not worth saving:

• Easily googleable facts
• Transient scratchpad work
• Anything the user will obviously remember themselves

Cleanup and Distillation

Do not aggressively reorganize or distill memories by default.

Only do memory cleanup when:

• The user asks to organize, clean up, archive, consolidate, or review memories
• You are already working inside the memory folder and obvious cleanup is part of the task

When cleaning up memories:

• Prefer marking outdated files with status: superseded or moving them into an archive/ folder if such a structure already exists or the user asks for it
• Avoid deleting memory files unless the user explicitly requests deletion
• Preserve historical context; do not collapse distinct events into one vague note
• Merge or simplify only when duplicates are clearly redundant and the result is more useful than the originals
• Keep tags consistent so search stays clean

Distill carefully:

• Promote repeated or clearly durable facts into semantic memory when you have good evidence they should be long-lived
• Promote repeatable workflows into procedural memory when they are clearly meant to be reused
• If you are not confident, preserve the original memory and avoid over-distilling

If the user asks to "clean up my memories", a good default is:

• review for stale or superseded files
• archive or mark outdated material
• tighten titles and tags
• leave meaningful history intact

When to Create vs Update a File

Create a new file when:

• You are recording a distinct event, session, decision point, incident, or milestone
• You are starting a new topic with no current memory for it
• You want to preserve a dated snapshot of what happened at a specific point in time

Update an existing file when:

• You are maintaining an ongoing semantic record of current state
• You are improving an existing procedure or checklist
• A stable file for this subject already exists and should remain the source of truth

The goal is to avoid both extremes: don't spam new files for every tiny change, but don't flatten meaningful history into one endlessly edited document either.

When in doubt:

• choose episodic if the value is historical context
• choose semantic if the value is current truth
• choose procedural if the value is repeatable guidance
• choose more than one if needed

When something evolves across clearly distinct phases, new files tell a useful story:

2025_03_09_supplier_negotiation_initial.md
2025_03_10_supplier_negotiation_counteroffer.md
2025_03_11_supplier_negotiation_final_terms.md

To mark a file as outdated, add status: superseded to its frontmatter — don't delete it.

File Organization

By default, organize memory by type:

• ./memory/episodic/
• ./memory/semantic/
• ./memory/procedural/

You may also encounter or be instructed to use a more structured layout such as ./memory/semantic/clients/ or ./memory/procedural/operations/. Follow whatever structure exists; if none exists, use the default type-based layout.