SharpAgent Layered Memory System v1.0.0

> Memory like a human — remembers what matters, forgets the rest, and searches accurately.

> Based on Layered Memory Sys (ClawHub, +324% in 24 days) 6-layer architecture + dream processing + Mem0 persistence.

Core Problem

Current agent memory problems:

SharpAgent memory solves all three with 6 layers + dream processing.

Memory Layers

                        [User interaction]
                            │
                   ┌────────┴────────┐
                   │   L1: Ephemeral │  ← Current session context
                   └────────┬────────┘
                            │ Session ends
                   ┌────────┴────────┐
                   │   L2: Short-Term│  ← Recent sessions (rolling window)
                   └────────┬────────┘
                            │ Periodic migration
                   ┌────────┴────────┐
                   │   L3: Working   │  ← Active task state
                   └────────┬────────┘
                            │ Task complete
                   ┌────────┴────────┐
                   │   L4: Contextual│  ← Task-level context
                   └────────┬────────┘
                            │ Dream processing
                   ┌────────┴────────┐
                   │   L5: Long-Term │  ← Preferences & lessons
                   └────────┬────────┘
                            │ Archive expired
                   ┌────────┴────────┐
                   │   L6: Archive   │  ← Historical archive
                   └─────────────────┘

L1: Ephemeral

Storage: Current session all messages

Capacity: Model context window

Lifespan: Session end = gone

Index: None, linear

Use: In-context understanding

L2: Short-Term

Storage: Last N sessions (default N=5)

Medium: In-memory (Redis optional)

Capacity: 5 session summaries, ≤2KB each

Lifespan: 7 days → auto-migrate to L4

Index: Session ID + timestamp + tags

Use: Quick cross-session reference

L3: Working

Storage: Active task state

Medium: JSON files (memory/working/)

Capacity: ≤10KB per task

Lifespan: Task complete → L4; task interrupted → kept

Index: Task ID + status + last update

Use: Resume interrupted tasks, multitasking

L4: Contextual

Storage: Completed task full context

Medium: JSON files (memory/contextual/) + optional SQLite

Capacity: Unlimited, but retrieval Top 5

Lifespan: Until dream processing (30d no reference → L5)

Index: TF-IDF full-text

Use: Look back at past tasks, reuse solutions

L5: Long-Term

Storage: Persistent cross-session knowledge

Medium: SQLite (memory/long_term.db) + Chinese tokenizer index

Capacity: Unlimited

Lifespan: Permanent unless explicitly forgotten

Index: TF-IDF + jieba Chinese tokenization

Use: User preferences, lessons, best practices, key decisions

L6: Archive

Storage: Expired or low-referenced L4/L5 entries

Medium: SQLite (memory/archive.db), read-only

Capacity: Theoretically infinite

Lifespan: Permanent read-only

Index: None (time + category)

Use: Legal compliance retention, audit trail

Dream Processing

Dreams aren't just for humans. Agents need low-load memory maintenance too.

Trigger: Heartbeat (low load, every 30 min), user says "clean up", or scheduled 04:00 daily.

Four Dream Operations:

1. Consolidation

Combine scattered memory fragments into coherent knowledge.

Input: Multiple fragments
→ "Use 150-char abstracts" (verified multiple times)
→ "User prefers shorter versions"
→ "Briefing read rate improved 30%"
→ Consolidate to:
  "Best practice: 150-char abstract in briefings (3x verified, +30% read rate)"

Trigger: Same pattern appears ≥3 times

2. Archiving

Move low-reference items out of working cache.

→ L4 entries with 0 references in 30 days
→ Move to L6 archive
→ Remove from L4 tag index

Trigger: 30-day reference count = 0

3. Forgetting

Actively delete low-value, duplicate, or outdated content.

→ "User once preferred Python 3.9" (3 months ago, now 3.13)
→ Outdated, delete
→ Keep space for valuable info

Trigger:

• Newer version available
• Explicitly contradicted
• >90 days with <2 references
• User says "forget this"

4. Merging

Combine multiple related L5 entries into higher-level patterns.

→ "Prefers 150-char abstract" (confidence=8)
→ "Prefers bullet points" (confidence=7)
→ "Dislikes tables" (confidence=6)
→ Merge:
  "User prefers briefings in bullet points + 150-char abstract, avoid tables"

Trigger: High confidence (≥7) + same category

Search

Standard Search

def search(query, layers=["L4", "L5"]):
    tokens = jieba.cut(query)      # Chinese tokenization
    vec = tfidf_vectorizer.transform(tokens)
    scores = cosine_similarity(vec, layer_index)
    return top_k(scores, k=5)

Chinese Tokenization

jieba.load_userdict("memory/custom_dict.txt")
# "五元组审查" → ["五元组", "审查"] not mis-split
# "惠迈校准框架" → ["惠迈", "校准", "框架"]

Search Priority

Storage Architecture

memory/
├── working/            # L3: Task working area (JSON)
├── contextual/         # L4: Contextual memory (JSON)
├── long_term.db        # L5: SQLite + FTS5 index
├── archive.db          # L6: Archive storage
├── custom_dict.txt     # Domain Chinese word dictionary
└── dream_log.json      # Dream processing log

Edge Cases

Quality Gates

Integration Points

Five-Factor Review

• Important memory entries verified before writing
• L5 entries carry FiveFactorResult

Self-Evolving Loop

• Reflections → L4 contextual, verified lessons → L5 long-term
• Dream consolidation output = new "best-practice" L5 entries

Engineering Lifecycle

• Working L3 supports task checkpoint/resume
• L4 stores engineering decision traces

Version History

• v1.0.0 — Initial release. 6-layer memory architecture with dream processing, TF-IDF search, Chinese tokenization.

SharpAgent · MIT-0 · 2026-05-11