Memory Systems & Learned Mental Lexicon

AI Friend implements a multi-tier cognitive memory architecture inspired by John R. Anderson's ACT-R Cognitive Architecture and semantic spreading activation networks.

---

The 4 Memory Tiers

1. Working Memory (Short-Term Buffer)

Maintains recent conversational turns, immediate sensory observations, and active conversational context in Redis and in-memory caches.

2. Episodic Memory (Lived History)

Stores conversation turns as 768-dimensional dense vectors in PostgreSQL (pgvector). Each memory records:

• content: The raw dialogue or observation.
• valence & arousal: Affective state at the time of encoding.
• recall_count: Total number of times recalled.
• last_recalled_at: Timestamp of latest activation.

3. Semantic Knowledge Graph (Neo4j)

A property graph mapping entities, relationships, shared preferences, and factual knowledge ((:User)-[:DISLIKES]->(:Topic)). Allows graph traversal and multi-hop reasoning.

4. Archival Tier

Memories whose activation score falls below the retention threshold move to cold archival storage rather than being deleted, preserving history for future long-term retrieval.

---

ACT-R Power-Law Memory Decay

Memories do not persist forever at full strength. Activation strength decays over time following the ACT-R power-law formula:

$A_i = \ln\left(\sum_{k=1}^{n} (t - t_k)^{-d}\right) + \sum_{j} W_j S_{ji}$

Where:

• $t - t_k$: Time elapsed since the $k$-th recall.
• $d$: Memory decay parameter (default $d = 0.5$).
• $S_{ji}$: Associative strength between cue $j$ and memory $i$.

Memories frequently discussed remain prominent; trivial one-off details naturally fade.

---

The Learned Mental Lexicon

Standard retrieval systems rely on generic static embeddings or rigid dictionary synonyms. AI Friend builds a dynamic mental lexicon extracted directly from past conversations:

• When you mention a nickname, a personal hobby, or an idiosyncratic phrase, the agent creates associative links in its lexicon graph.
• When searching memories for "weekend project", the lexicon expands the search query using terms unique to your relationship (e.g. "compiler", "rust", "garage").

---

Subconscious REM Sleep Consolidation

When the conversational mesh is idle, subconscious_agent triggers an offline consolidation cycle:

1. Fact Extraction: Extracts new permanent facts and beliefs from recent turns.
2. Knowledge Graph Graph-Sync: Updates Neo4j entity relationships.
3. Decay Refresh: Recalculates ACT-R activation scores across all stored memories.
4. Proactive Reflection: Generates unprompted spontaneous thoughts that can trigger proactive outreach.