Classification System

abmind uses a NATO Admiralty Code-inspired system to rate memory reliability and control access. Every extracted memory carries three quality scores and a classification level.

Classification Levels

Class	Name	Access	Encryption	Example
0	Public	All roles	None	General facts, preferences
1	Internal	user + master	None	Personal details, work context
2	Confidential	user + master	None	Sensitive decisions, private opinions
3	Secret	master only	AES-256-GCM at rest	API keys, passwords, credentials

Default classification for new memories: 1 (Internal).

Access control

Recall queries include maxClassification (derived from the user's role):

• master → sees class 0–3
• user → sees class 0–2
• guest → sees class 0 only

Class 3 memories are encrypted in the database. They're decrypted only when recalled by a master-role user, and only via explicit tool call (never auto-injected into context).

Quality Scores

Trust (source reliability)

How reliable is the source of this information?

Score	Meaning	Assigned when
0	Unknown	Default for new memories
1	Unreliable	Contradicted by other evidence
2	Fairly reliable	Single unconfirmed source
3	Reliable	Confirmed by user or multiple sources

Integrity (information accuracy)

How accurate is the information itself, independent of source?

Score	Meaning	Assigned when
0	Unknown	Default
1	Doubtful	Contradicts known facts
2	Possibly true	Plausible but unconfirmed
3	Probably true	Consistent with other memories
4	Confirmed	Explicitly verified by user

Default: 2 (Possibly true).

Credibility (combined assessment)

Composite score combining trust and integrity. Used as a ranking multiplier in recall:

Score	Effect on recall ranking
≤ 2	1.25× boost (high credibility)
3–4	1.0× (neutral)
≥ 5	1.0× (neutral — no penalty, just no boost)

Lower credibility number = more credible. Scale: 0 (highest) to 6 (lowest).

Default: 6 (unassessed).

How scores affect recall

The Sf stage scoring formula includes:

trust_factor = 0.5 + 0.5 × (trust / 3)
credibility_factor = 1.25 if credibility ≤ 2, else 1.0

A memory with trust=3, credibility=1 gets a 1.0 × 1.25 = 1.25× boost over an unscored memory (trust=0, credibility=6 → 0.5 × 1.0 = 0.5×).

Setting scores

Automatic (during extraction)

The sleep cycle's extraction step assigns initial scores based on:

• Source: user-stated facts get higher trust than inferred ones
• Confidence: LLM extraction confidence maps to integrity
• Type: secrets auto-classify as class 3

Manual (via edit)

abmind edit --id 42 --trust 3 --integrity 4 --credibility 1 --classification 2

Via tool call (agent-initiated)

The memory_store and memory_edit tools accept trust, integrity, credibility, and classification parameters.

Multi-user isolation

Classification enforcement is per-query, not per-memory. The same memory can be visible to one user and hidden from another based on their maxClass setting.

{
  "users": [
    { "userId": "admin", "role": "master", "maxClass": 3 },
    { "userId": "colleague", "role": "user", "maxClass": 1 }
  ]
}

In this setup, colleague sees only class 0–1 memories. Class 2–3 memories exist in the same database but are filtered out at query time.

Encryption (Class 3)

Class 3 memories use AES-256-GCM encryption:

• Key stored at ~/.abmind/secret/abmind.key (chmod 600)
• Each memory gets a unique IV
• Encrypted content stored in content_en column (base64-encoded ciphertext)
• content_original also encrypted if present

Commands:

abmind encrypt-secrets    # encrypt existing class-3 rows
abmind list-secrets       # show metadata (no decryption)
abmind rekey --old-key /path/to/old.key   # re-encrypt with new key

Upgrading classification

Promote a memory to higher classification:

abmind edit --id 42 --classification 3
abmind encrypt-secrets    # encrypt the newly-classified row

Demoting from class 3 requires --user-override flag (safety gate):

abmind edit --id 42 --classification 1 --user-override