making human memory.

Built upon decades of cognitive science research on memory storage & retrieval.

Three-Tier Memory Model

Based on Atkinson-Shiffrin's multi-store model [1] and Tulving's distinction between episodic and semantic memory [3]. Information flows from immediate sensory processing through interpretation to long-term consolidation.

WORKING MEMORY Baddeley's central executive model PROPERTIES capacity: 7±2 chunks (Miller) duration: ~20 seconds decay: rapid without rehearsal function: active processing "just mentioned coffee" "current topic: work stress" "asked about weekend" consolidation SEMANTIC MEMORY Tulving's semantic/episodic distinction PROPERTIES capacity: ~1000 active items duration: decay-dependent index: HNSW vector search function: interpreted meaning "prefers mornings" "works in software" "has dog named Max" archive LONG-TERM ARCHIVE core identity, permanent storage PROPERTIES capacity: unlimited duration: permanent index: HNSW + BM25 hybrid function: identity persistence "introvert, values depth" "birthday: March 15" "core value: authenticity"

Gist-Based Extraction

Following Roediger & Karpicke's research on gist memory [4], Soul extracts semantic meaning rather than storing verbatim transcripts. Fuzzy-trace theory [5] shows humans primarily retain interpreted gist, not surface details.

VERBATIM INPUT "I usually wake up around 6am because I like having quiet time before work" LLM extraction GIST MEMORY "early riser, values quiet/solitude" saliency: 4 | epistemic: known Roediger & Karpicke (2006)

Retrieval-Based Strengthening

Roediger & Karpicke's testing effect [4]: memories strengthen through retrieval, not just exposure. Each time Soul recalls a memory in context, its retrieval strength increases and decay rate decreases.

Query current context Embed q ∈ ℝ¹⁵³⁶ VECTOR SPACE q RETRIEVAL EFFECT on each retrieval: accessCount++ lastAccessed = now() decayRate *= 0.95 ↑ retrieval strength

Metacognitive Confidence

Based on Koriat's research on Feeling of Knowing (FOK) and Judgments of Learning (JOL) [6]. Soul tracks epistemic states—distinguishing facts from inferences, certainty from speculation.

KNOWN high FOK, explicit statement "birthday is March 15" "works at Acme Corp" "has two children" HYPOTHESIS lower FOK, inferred "may be career change?" "possibly stressed lately" "might prefer texts to calls" Koriat (1993, 2007)

Saliency-Weighted Decay

Ebbinghaus's forgetting curve [7] combined with importance-weighting. High-saliency memories resist decay; low-saliency memories fade rapidly. Retrieval slows decay rate [8].

time since encoding retention sal=5 sal=4 sal=3 sal=2 sal=1 deletion threshold retention(t) = e^(-λ[s]·t) × (1 + log(retrievals + 1)) Ebbinghaus (1885), Bahrick (1984)

References

  1. Atkinson, R.C. & Shiffrin, R.M. (1968). Human memory: A proposed system and its control processes.
  2. Baddeley, A. (2000). The episodic buffer: a new component of working memory?
  3. Tulving, E. (1972). Episodic and semantic memory.
  4. Roediger, H.L. & Karpicke, J.D. (2006). The power of testing memory.
  5. Brainerd, C.J. & Reyna, V.F. (2002). Fuzzy-trace theory and false memory.
  6. Koriat, A. (1993). How do we know that we know? The accessibility model of feeling of knowing.
  7. Ebbinghaus, H. (1885). Memory: A contribution to experimental psychology.
  8. Bahrick, H.P. (1984). Semantic memory content in permastore.