memorydrift 3.0.2

AI memory that decays, degrades, and drifts the way human memory does

MemoryLayer

AI memory that decays, degrades, and drifts the way human memory does.

Install

pip install memorydrift
pip install "memorydrift[embed]"
pip install "memorydrift[full]"

Quickstart

from memorylayer import MemoryStore, MemoryLayer

store = MemoryStore(auto_embed=True)

store.write("I saved the blacksmith's son", emotional_weight=0.9, age_at_encoding=28)
store.write("The dragon was defeated in the northern tower", emotional_weight=0.7)

for r in store.retrieve():
    print(r.display_content)   # "I clearly remember: I saved the blacksmith's son"
    print(r.decay.fidelity)    # VIVID / CLEAR / FADED / VAGUE / FEELING

Spreading activation fires automatically — semantically similar memories surface together.

What makes it different

Every other AI memory system is a key-value store. MemoryLayer implements the neuroscience:

Human memory property	MemoryLayer
Emotional memories last longer	emotional_weight stretches halflife up to 3×
Details fade before the gist	Fidelity degrades: VIVID → CLEAR → FADED → VAGUE → FEELING
Rehearsal keeps memories strong	repetition_count boosts stability (spaced repetition)
Related memories surface together	Spreading activation via semantic embeddings
Weak memories drift on recall	Reconsolidation: content rewrites toward current context
Childhood memories are hazy	Age-at-encoding: childhood amnesia + reminiscence bump (Rubin 1997)
Sleep consolidates the day	simulate_sleep() adds replay repetitions, tags CLEAR+ for promotion

Memory layers

Layer	Halflife	Use for
WORKING	~1 hour	Active conversation context
EPISODIC	1 year	Events and experiences
SEMANTIC	5 years	Facts and knowledge
PROSPECTIVE	3 days	Intentions that fade if unacted on
IDENTITY	Forever	Core beliefs and self-concept

Advanced API

from memorylayer import (
    # Prospective memory — intentions
    write_intention, complete_intention, get_overdue_intentions,

    # Collective memory — shared events, per-entity emotional weight
    CollectiveMemoryStore,

    # Sleep consolidation
    simulate_sleep, SleepSession,

    # Reconsolidation — memory drift on recall
    reconsolidate_sync,
)

# Prospective memory
store = MemoryStore()
intention = write_intention(store, "Call the lawyer", due_in_days=3)
complete_intention(store, intention.id)   # converts to episodic on completion

# Collective memory
shared = CollectiveMemoryStore()
event = shared.write_event(
    "The team won the championship",
    participants={"alice": 0.95, "bob": 0.40},   # alice scored; bob watched
)
alice_memories = shared.recall_for("alice")       # stronger than bob's

# Sleep consolidation
result = simulate_sleep(store, SleepSession(quality=0.85, duration_hours=7.5))
print(f"Strengthened: {result.memories_strengthened}")

Fidelity levels

When a memory is retrieved, its content is automatically degraded to match its strength:

VIVID    (> 0.70)  "I clearly remember: she smiled at the coffee shop."
CLEAR    (0.45–)   "I remember: something about a coffee shop."
FADED    (0.22–)   "I vaguely recall: she smiled."
VAGUE    (0.08–)   "I have a faint sense that: she smiled…"
FEELING  (0.03–)   "I don't remember what happened, but I remember warmth and joy."
FORGOTTEN(< 0.03)  [not returned — the memory is gone]

Patent

MemoryLayer implements three novel mechanisms filed with the Indian Patent Office (June 2026):

1. Emotionally-modulated decay floor — high-emotion memories asymptotically approach a floor rather than decaying to zero.
2. Fidelity-degraded retrieval — content specificity decreases proportionally to computed temporal strength.
3. Post-retrieval reconsolidation — weak memories are rewritten toward current context on each recall.

---

Author: Mounica Goriparti · [e-mail]