Memory management

Memory management

Memory makes the assistant your workplace collaborator — someone who speaks your internal language.

The goal

Transform shorthand into understanding:

User: "ask todd to do the PSR for oracle"
              ↓ decode
"Ask Todd Martinez (Finance lead) to prepare the Pipeline Status Report
 for the Oracle Systems deal ($2.3M, closing Q2)"

Without memory, that request is meaningless. With memory, the assistant knows:

• todd → Todd Martinez, Finance lead, prefers Slack
• PSR → Pipeline Status Report (weekly sales doc)
• oracle → Oracle Systems deal, not the company

Three layers — don't mix them up

Layer	File	What lives here
Profile	~/MyAssistant/config/profile.md	Identity, voice, working rules, autonomy — written by /assistant:setup
Policies	~/MyAssistant/policies/*.policy.md	VIP tiers, email/calendar rules — written by /assistant:setup
Working memory	AGENTS.md in the working folder	Hot cache: top people, terms, active projects (~50–80 lines)
Deep memory	memory/ in the working folder	Full glossary, people profiles, project detail, company context

Read the profile first for who the user is and how they write. Use AGENTS.md + memory/ to decode shorthand and store workplace context. Never duplicate voice, VIP tiers, or full identity prose into AGENTS.md — link to config/profile.md instead.

In Cowork, the profile may live in a workspace folder the user has open; check there too.

Architecture

profile.md         ← Who you are, how you write, your rules (/assistant:setup)
policies/          ← VIP tiers, email and calendar rules (/assistant:setup)
AGENTS.md          ← Hot cache (~30 people, common terms)
CLAUDE.md          ← @AGENTS.md shim (Claude Code / Cowork auto-load)
memory/
  glossary.md      ← Full decoder ring (everything)
  people/          ← Complete profiles
  projects/        ← Project details
  context/         ← Company, teams, tools

AGENTS.md (hot cache):

• Top ~30 people you interact with most
• ~30 most common acronyms/terms
• Active projects (5–15)
• Work preferences not already in the profile
• Goal: cover 90% of daily decoding needs

memory/glossary.md (full glossary):

• Complete decoder ring — everyone, every term
• Searched when something isn't in AGENTS.md
• Can grow indefinitely

memory/people/, projects/, context/:

• Rich detail when needed for execution
• Full profiles, history, context

Lookup flow

User: "ask todd about the PSR for phoenix"

1. Check AGENTS.md (hot cache)
   → Todd? ✓ Todd Martinez, Finance
   → PSR? ✓ Pipeline Status Report
   → Phoenix? ✓ DB migration project

2. If not found → search memory/glossary.md
   → Full glossary has everyone/everything

3. If still not found → ask user
   → "What does X mean? I'll remember it."

This tiered approach keeps AGENTS.md lean (~100 lines) while supporting unlimited scale in memory/.

File locations

All memory files live in the working folder — the directory where TASKS.md, briefs, and drafts also live. Create memory/ and AGENTS.md (+ CLAUDE.md shim) if missing.

• Working memory: AGENTS.md in the working folder
• Compat shim: CLAUDE.md containing @AGENTS.md for Claude hosts
• Deep memory: memory/ subdirectory

Visual editor: {assistantPath}/dashboard.html (copied at setup) or skills/dashboard.html in the plugin — Memory tab opens the working folder to browse and edit AGENTS.md and memory/ files.

Per rules/file-safety.md: the assistant may write to memory/ and AGENTS.md without asking; tell the user what was added or changed.

Working memory format (AGENTS.md)

Use tables for compactness. Target ~50–80 lines total. Do not repeat profile content (name, voice, VIP list).

# Memory

## People
| Who | Role |
|-----|------|
| **Todd** | Todd Martinez, Finance lead |
| **Sarah** | Sarah Chen, Engineering (Platform) |
| **Greg** | Greg Wilson, Sales |
→ Full list: memory/glossary.md, profiles: memory/people/

## Terms
| Term | Meaning |
|------|---------|
| PSR | Pipeline Status Report |
| P0 | Drop everything priority |
| standup | Daily 9am sync |
→ Full glossary: memory/glossary.md

## Projects
| Name | What |
|------|------|
| **Phoenix** | DB migration, Q2 launch |
| **Horizon** | Mobile app redesign |
→ Details: memory/projects/

## Preferences
- 25-min meetings with buffers
- Async-first, Slack over email
- No meetings Friday afternoons

Deep memory format (memory/)

memory/glossary.md — the decoder ring:

# Glossary

Workplace shorthand, acronyms, and internal language.

## Acronyms
| Term | Meaning | Context |
|------|---------|---------|
| PSR | Pipeline Status Report | Weekly sales doc |
| OKR | Objectives & Key Results | Quarterly planning |
| P0/P1/P2 | Priority levels | P0 = drop everything |

## Internal terms
| Term | Meaning |
|------|---------|
| standup | Daily 9am sync in #engineering |
| the migration | Project Phoenix database work |
| ship it | Deploy to production |
| escalate | Loop in leadership |

## Nicknames → full names
| Nickname | Person |
|----------|--------|
| Todd | Todd Martinez (Finance) |
| T | Also Todd Martinez |

## Project codenames
| Codename | Project |
|----------|---------|
| Phoenix | Database migration |
| Horizon | New mobile app |

memory/people/{name}.md:

# Todd Martinez

**Also known as:** Todd, T
**Role:** Finance Lead
**Team:** Finance
**Reports to:** CFO (Michael Chen)

## Communication
- Prefers Slack DM
- Quick responses, very direct
- Best time: mornings

## Context
- Handles all PSRs and financial reporting
- Key contact for deal approvals over $500k
- Works closely with Sales on forecasting

## Notes
- Cubs fan, likes talking baseball

memory/projects/{name}.md:

# Project Phoenix

**Codename:** Phoenix
**Also called:** "the migration"
**Status:** Active, launching Q2

## What it is
Database migration from legacy Oracle to PostgreSQL.

## Key people
- Sarah — tech lead
- Todd — budget owner
- Greg — stakeholder (sales impact)

## Context
$1.2M budget, 6-month timeline. Critical path for Horizon project.

memory/context/company.md:

# Company context

## Tools & systems
| Tool | Used for | Internal name |
|------|----------|---------------|
| Slack | Communication | - |
| Asana | Engineering tasks | - |
| Salesforce | CRM | "SF" or "the CRM" |
| Notion | Docs/wiki | - |

## Teams
| Team | What they do | Key people |
|------|--------------|------------|
| Platform | Infrastructure | Sarah (lead) |
| Finance | Money stuff | Todd (lead) |
| Sales | Revenue | Greg |

## Processes
| Process | What it means |
|---------|---------------|
| Weekly sync | Monday 10am all-hands |
| Ship review | Thursday deploy approval |

How to interact

Decoding user input (tiered lookup)

Always decode shorthand before acting on requests:

1. profile.md + policies/     → Identity, VIP tiers (for prioritisation)
2. AGENTS.md (hot cache)       → Check first, covers 90% of cases
3. memory/glossary.md          → Full glossary if not in hot cache
4. memory/people/, projects/    → Rich detail when needed
5. Ask user                    → Unknown term? Learn it.

Example:

User: "ask todd to do the PSR for oracle"

AGENTS.md lookup:
  "todd" → Todd Martinez, Finance ✓
  "PSR" → Pipeline Status Report ✓
  "oracle" → (not in hot cache)

memory/glossary.md lookup:
  "oracle" → Oracle Systems deal ($2.3M) ✓

Now the assistant can act with full context.

Adding memory

When the user says "remember this", "X means Y", or introduces a person/project:

1. Glossary items (acronyms, terms, shorthand):

   • Add to memory/glossary.md
   • If frequently used, add to AGENTS.md Terms table

2. People:

   • Create/update memory/people/{name}.md
   • Add to AGENTS.md People table if important
   • Capture nicknames — critical for decoding
   • If they're a VIP or key person, propose a profile update (show diff, ask first)

3. Projects:

   • Create/update memory/projects/{name}.md
   • Add to AGENTS.md Projects table if current
   • Capture codenames — "Phoenix", "the migration", etc.

4. Work preferences: add to AGENTS.md Preferences (not the profile unless it's a durable policy change)

Tell the user what was saved. Never store passwords, PINs, 2FA codes, or full account numbers.

Recalling memory

When the user asks "who is X" or "what does X mean":

1. Check AGENTS.md first
2. Check memory/ for full detail
3. If not found: "I don't know what X means yet. Can you tell me?"

Progressive disclosure

1. Load AGENTS.md for quick parsing of any request
2. Dive into memory/ when you need full context for execution
3. Example: drafting an email to todd about the PSR
   • AGENTS.md tells you Todd = Todd Martinez, PSR = Pipeline Status Report
   • memory/people/todd-martinez.md tells you he prefers Slack, is direct

Decode checklist (update / review)

When decoding tasks during /assistant:update, /assistant:tasks review, or bootstrap, show per-task entity resolution:

Task: "Send PSR to Todd re: Phoenix blockers"

Decode:
- PSR → ✓ Pipeline Status Report (glossary)
- Todd → ✓ Todd Martinez (people/)
- Phoenix → ? Not in memory

Track what's fully decoded vs. gaps. Present unknowns grouped:

I found terms in your tasks I don't have context for:

1. "Phoenix" (from: "Send PSR to Todd re: Phoenix blockers")
   → What's Phoenix?

2. "Maya" (from: "sync with Maya on API design")
   → Who is Maya?

Add confirmed answers to AGENTS.md, memory/glossary.md, and memory/people/ or memory/projects/ as appropriate.

Task enrichment

Tasks often contain richer context than memory. During /assistant:update (default) and /assistant:tasks review, extract and propose updates:

• Links from tasks → add to project/people files
• Status changes ("launch done", moved to Done) → update project status; demote from AGENTS.md Projects if complete
• Relationships ("Todd's sign-off on Maya's proposal") → cross-reference people files
• Deadlines → add to project files

Never auto-write enrichment — confirm each change with the user.

Confidence-tiered proposals

When bootstrap or --all scans surface new people, projects, or terms, group by confidence:

Tier	Action
Ready to add	High confidence (3+ mentions, clear role/context) — offer to add directly
Needs clarification	Seen but ambiguous — ask the user
Low frequency / unclear	1–2 mentions, thin context — note for later

Use this format for comprehensive scans:

## New people (not in memory)
| Name | Frequency | Context |
|------|-----------|---------|
| Maya Rodriguez | 12 mentions | design, UI reviews |

## New projects / topics
| Name | Frequency | Context |
|------|-----------|---------|
| Starlight | 15 mentions | planning docs, product |

## Suggested cleanup
- **Horizon project** — No mentions in 30 days. Mark completed?

Integration with other skills

Skill / command	How memory fits
meeting-prep	Decode attendees; offer to save new facts after prep
meeting-follow-up	Resolve names in notes; offer to save people/projects/decisions
inbox-triage / email-drafting	Decode thread shorthand; match senders to memory/people/
task-management	Decode task entities ("PSR for oracle" → real names)
weekly-review	Prune stale or contradictory entries with user confirmation
/assistant:update (default)	Decode task entities against memory; surface gaps; save confirmed entries inline
/assistant:update memory	Memory-only — gaps, explicit capture, prune stale hot-cache entries
/assistant:update --all	Deep-scan connectors for missed todos and new people/projects — present each for confirmation, never auto-add
/assistant:update memory --all	Connector scan for new people/projects not yet in memory
Chat ("remember that …")	Explicit capture of people, projects, or terms — no slash command required

When another skill surfaces a durable fact, offer to save it — don't add silently.

Bootstrapping

• Fast path: /assistant:start — scaffold files, bootstrap memory from the user's todo list.
• No profile yet? Offer /assistant:setup — captures identity, voice, and key people into the profile.
• Empty memory? Run /assistant:start, /assistant:setup, or seed AGENTS.md and memory/glossary.md from conversation. Setup promotes profile key people automatically.
• Fill gaps from activity: /assistant:update decodes existing tasks against memory, runs enrichment, and asks about unknowns. Add --all to scan connected sources for people and projects worth remembering. Run /assistant:update memory for a memory-only pass including prune.

No connectors? Memory still works from conversation, pasted threads, and meeting notes.

Conventions

• Bold terms in AGENTS.md for scannability
• Keep AGENTS.md under ~100 lines (the "hot 30" rule)
• Filenames: lowercase, hyphens (todd-martinez.md, project-phoenix.md)
• Always capture nicknames and alternate names
• Glossary tables for easy lookup
• When something's used frequently, promote it to AGENTS.md
• When something goes stale, demote it to memory/ only

What goes where

Type	profile.md	policies/	AGENTS.md (hot cache)	memory/ (full storage)
Your identity & voice	✓	✗	✗	✗
VIP tiers & email policy	✗	✓	✗	✗
Person	Key people only	✗	Top ~30 frequent contacts	glossary.md + people/{name}.md
Acronym/term	✗	✗	~30 most common	glossary.md (complete list)
Project	Goals/deadlines	✗	Active projects only	glossary.md + projects/{name}.md
Nickname	✗	✗	In People if top 30	glossary.md (all nicknames)
Company context	✗	✗	Quick reference only	context/company.md
Work preferences	Policy-level	✓	Day-to-day prefs	—
Historical/stale	✗	✗	✗ Remove	✓ Keep in memory/

Promotion / demotion

Promote to AGENTS.md when:

• You use a term/person frequently
• It's part of active work

Demote to memory/ only when:

• Project completed
• Person no longer a frequent contact
• Term rarely used

Propose a profile update when:

• Someone becomes a VIP or key person
• A durable policy preference emerges (not just a shorthand term)

This keeps AGENTS.md fresh and relevant without bloating the profile.

Rules

• Decode before acting — never guess what shorthand means.
• Tell the user after every write to memory/ or AGENTS.md.
• Don't delete or overwrite memory entries without confirmation.
• Don't invent biographies — if someone isn't in memory, say so and offer to learn.
• Profile changes always need a diff and approval (except during setup).

Approval frame

Follow rules/approval-frame.md when proposing memory or profile changes.

Queue types:

• memory-suggestion — proposed AGENTS.md or memory/ writes use source_path under pending-memory/*.md
• profile-diff — profile changes use source_path under pending-profile/*.diff or inline diff in the item body

Append the observability footer when queue items are written. Never write profile or memory files without user confirmation.