Memory System

Deprex’s memory system enables the FDE to autonomously accumulate, reference, and organize organizational context. PMs do not need to create memory manually. The FDE learns through workflow observation and retains necessary knowledge as memory.

Memory is Autonomously Accumulated

Conventional knowledge bases follow a “humans write documents, AI searches them” model. Deprex’s memory system inverts this — the FDE itself extracts knowledge from the following sources:

• Slack conversations: Discussions, decisions, and business patterns within channels
• Meeting transcripts: Transcriptions and minutes from tools like Tldv
• Shared documents: Materials shared via Google Drive, Notion, etc.
• Task execution results: Feedback gained from the FDE’s own work
• PM interactions: Business knowledge implicitly shared through Slack conversations

The FDE automatically saves observed information as memory files and auto-updates INDEX.md (the memory catalog).

3-Layer Memory Structure

Memory is managed in three layers. Each layer is independent, and the FDE references the appropriate layer based on task context.

Layer	Location	Content	Management
Shared	.memory/	Knowledge common across all repositories	Auto-accumulated
Repo-specific	.deprex/brain/memory/	Repository-specific knowledge	Auto-accumulated
Tenant-specific	app/orgs/<company>/memory/	Tenant-specific business knowledge	Auto-accumulated + PM review

PMs primarily review tenant-specific memory. Shared and repo-specific memory are autonomously managed by the FDE.

Automatic Recall

Before executing a task, the FDE automatically searches for and references relevant memory.

Request: "Create this week's KPI report"

FDE's internal process:
1. Search INDEX.md recall_cues
2. Reference kpi-definitions.md -> retrieve KPI definitions and calculation methods
3. Reference reporting-rules.md -> retrieve report format and notification targets
4. Reference stakeholders.md -> identify reviewers
5. Generate the report and post to Slack

PMs do not need to instruct “read this memory.” The FDE autonomously determines which memory is needed based on the request.

INDEX.md — The Memory Catalog

Each memory directory contains an INDEX.md. This is a catalog of memory files containing recall_cues that help the FDE efficiently find needed memory.

# Acme Memory Index

## kpi-definitions.md
- recall_cue: When you need to check Acme's KPI names, calculation methods, and week-over-week interpretation

## reporting-rules.md
- recall_cue: When you need to check standard report format, notification targets, and escalation conditions

## stakeholders.md
- recall_cue: When you need to check Acme's PM, engineering lead, and approvers

INDEX.md is auto-updated every time the FDE adds memory. PMs do not need to maintain it manually.

Automatic Distill (Organization)

Memory is not just accumulated. The FDE periodically runs automatic organization (Memory Distill):

• Staleness detection: Detects outdated information and proposes updates or removal
• Duplicate consolidation: Proposes merging similar memory files to prevent information scattering
• recall_cue optimization: Improves recall_cues to increase search accuracy

Distill results are notified to the PM. PMs simply respond with “this consolidation is correct” or “this information is still valid.”

PM Involvement: Review and Approval

PMs are not in the position of creating memory, but of reviewing what the FDE has learned.

Example Notification from FDE to PM

[Memory Update] Updated stakeholders.md
Change: Updated Tanaka-san's role from "Sales Manager" to "Division Director"
Source: 5/1 announcement in #general

PMs can take the following actions:

• Approve: If the change is correct, no action needed (already auto-applied)
• Correct: If there’s an error, point it out in Slack and the FDE will fix it
• Request deletion: Ask the FDE to remove unnecessary memory

Information Suited for Memory

Examples of information the FDE automatically accumulates as memory:

• Organizational structure and team member roles
• KPI definitions, calculation methods, and review cadence
• Report formats and notification targets
• Expressions to avoid in customer communications
• Workflow procedures and decision criteria

Information Not Stored as Memory

The FDE does not store the following as memory, instead retrieving them from source systems when needed:

• Daily fluctuating numerical data (KPI actuals, inventory counts, etc.)
• API keys, passwords, and other credentials (managed in config)
• Temporary task lists (managed in task management tools)

Cross-Session Memory Sharing

Memory is managed via git, so it is shared across multiple machines and sessions. Knowledge the FDE learned from a Slack conversation one day is available during scheduled task execution the next day.

Memory quality directly impacts FDE output quality. While the FDE autonomously maintains memory quality, periodic review by PMs further enhances it.