Memory Strata — 四层分层记忆架构

Four-layer stratified memory architecture for long-horizon AI agent projects. Solves critical fact forgetting, memory bloat, and knowledge sedimentation failures.

Architecture

L1  System Injection   MEMORY.md — key points + anchors        (auto-loaded every session)
L2  Project Core       project-core.md — stable facts + ADR     (rarely changes)
L3  Dev Status         dev-status.md — progress/params/issues     (updates with progress)
L4  Daily Journal      YYYY-MM-DD.md — session log              (per-session)

Three Mechanisms

1. Auto-Distillation (Episodic → Semantic)

• L4→L3: milestone completed, or referenced across ≥2 sessions
• L3→L2: stable ≥1 month + confirmed fact + likely needed in future
• L2→L1: useful beyond this project (user preferences, general lessons)

2. Memory Strength Decay (last_used)

• Tag entries with [last_used: YYYY-MM-DD]

• 14 days unused → ⚡ rarely used

• 30 days unused → auto-archive

3. Zettelkasten Link Enhancement

• Bidirectional [[...]] links between L2↔L3
• Check link integrity with the distillation script

Quick Start

1. Initialize a project

npx pi-memory-strata init-project <project-name>

2. Run daily maintenance (or schedule with cron)

npx pi-memory-strata maintain --vault /path/to/vault

3. Check memory health

npx pi-memory-strata check --vault /path/to/vault

Installation

As Pi Extension

pi install npm:pi-memory-strata

As npm package

npm install -g pi-memory-strata

Direct usage (no install)

npx pi-memory-strata <command>

Programmatic API

import { MemoryStrata } from 'pi-memory-strata';

const strata = new MemoryStrata({
  vaultPath: '/path/to/obsidian-vault',
  autoArchiveDays: 30,
  staleWarningDays: 14,
});

// Run distillation check
const check = await strata.check();
console.log(check.ok);

// Run daily maintenance
const report = await strata.maintain();
console.log(report.health);

// Initialize new project
strata.initProject('my-project');

Vault Structure

obsidian-vault/
├── 00-brain/
│   └── MEMORY.md              # L1: System injection
├── 10-journal/
│   ├── maintenance/           # Daily maintenance reports
│   ├── summary/               # Auto-generated daily summaries
│   └── 2026-01-01.md        # L4: Daily logs
├── 20-projects/
│   └── my-project/
│       ├── dev-status.md    # L3: Development status
│       └── project-core.md  # L2: Project core archive
├── 30-knowledge/
├── 60-learnings/
│   └── archive/             # Auto-archived entries
├── scripts/
│   ├── memory_distill_check.py
│   └── daily_memory_maintenance.py
└── templates/
    ├── 01-status.md
    └── 02-core.md

Daily Maintenance (Automated)

Schedule daily maintenance with cron or systemd:

# Every day at 09:30
30 9 * * * npx pi-memory-strata maintain

Maintenance performs:

1. Journal coverage check — verify last 7 days have logs
2. Dev status freshness — warn if >7 days without update
3. Core archive freshness — warn if >30 days without update
4. Memory decay — identify entries with stale last_used
5. Anchor integrity — verify memory anchors exist
6. Bidirectional links — check L2↔L3 link completeness
7. Auto-archive — move >30d stale entries to 60-learnings/archive/
8. Daily summary — generate previous day's summary

Academic Context

| Concept | Source | This Package | |---------|--------|-------------| | Write-Manage-Read loop | arXiv:2603.07670 | W(journal) + M(distill/decay) + R(search) | | Episodic→Semantic promotion | A-MEM (NeurIPS 2025) | L4→L3→L2 distillation | | Memory strength decay | MemoryBank (2024) | last_used + 30-day archive | | Zettelkasten linking | A-MEM | Obsidian [[...]] bidirectional links | | Atomic facts | AtomMem (2026) | Memory anchors with consequence tags |

License

MIT