neuram

Memory for AI agents that works the way your brain does. Not a vector database with a fancy name — actual neuroscience, implemented.

neuram (neural + RAM) uses real brain regions as first-class Python objects. Your agent has a Hippocampus. It has a PrefrontalCortex that holds 7 things at once, just like yours. Memories decay when ignored and survive when revisited. At idle, a sleep cycle quietly consolidates what matters and prunes what doesn't.

If you've ever built an agent and wondered why it keeps "forgetting" things it should know, or why it treats a one-off fact the same as something the user says every single day — this is the missing piece.

---

How it works

SensoryCortex → Thalamus → Amygdala → PrefrontalCortex
                                             ↓
                                       Hippocampus
                                             ↓  (sleep cycle)
                Neocortex / Cerebellum / BasalGanglia  (Redis)

Everything lives in-memory until it earns a spot in Redis. The same way your brain doesn't bother encoding every conversation into long-term storage — only the things that matter make it through.

Region	What it does	Where
SensoryCortex	Raw input buffer. Expires in 3 seconds if nobody pays attention.	In-memory
Thalamus	The bouncer. Filters out low-relevance input before it reaches working memory.	In-memory
Amygdala	Tags emotional weight. High-salience memories get a stability boost right at encoding.	In-memory
PrefrontalCortex	Working memory. 7±2 slots, LRU eviction, decays without rehearsal.	In-memory
Hippocampus	Episodic index. Handles pattern completion and temporal association.	In-memory
Neocortex	Long-term semantic and episodic storage.	Redis
Cerebellum	Procedural memory — learned skills and workflows. Very hard to unlearn once consolidated.	Redis
BasalGanglia	Habit memory, modulated by a dopamine-style reward signal.	Redis
SleepCycle	Background consolidation. Replays working memory, promotes to LTM, prunes the rest.	—

---

The memory unit: Engram

Richard Semon coined "engram" in 1904 to describe the physical trace a memory leaves in the brain. That's what we store.

@dataclass
class Engram:
    content: str
    memory_type: MemoryType     # EPISODIC | SEMANTIC | PROCEDURAL | HABITUAL
    salience: float             # 0–1, set by the Amygdala
    embedding: list[float]      # sentence-transformer vector
    synaptic_strength: float    # current activation — decays via Ebbinghaus
    stability: float            # resistance to forgetting — grows with each recall
    access_count: int
    associations: list[tuple]   # spreading activation graph edges
    fragments: list[str]        # sub-engram IDs — memory is reconstructive, not holistic

Two properties matter most: synaptic_strength (how vivid the memory is right now) and stability (how long it'll last before fading). Stability grows every time the memory is recalled — that's Long-Term Potentiation.

---

The biology, briefly

Forgetting curve. Ebbinghaus figured this out in 1885 by memorizing nonsense syllables and testing himself obsessively. The decay follows:

R(t) = e^(-t / S)

t is hours since last access. S is stability — which grows with each recall. So the more often you retrieve something, the longer it survives between retrievals. Spaced repetition falls out naturally from the math.

Long-Term Potentiation. Every recall event strengthens the engram:

stability += 0.3 × salience_boost × log(1 + access_count)

High-salience memories — the ones the Amygdala flagged as emotionally significant — grow faster. That's the norepinephrine effect.

Spreading activation. When you recall something, you don't just retrieve that one thing — you light up a neighborhood. A memory of "Paris" activates "France", "Eiffel Tower", "that trip in 2019". neuram does the same: cosine similarity gets initial activation, then it propagates through the association graph with a 0.6x decay per hop.

Sleep cycle. This is the one people don't build. While your agent is idle, a background coroutine runs: it replays working memory, promotes anything that earned it to long-term storage, applies Ebbinghaus decay to everything in Redis, and deletes engrams that fell below 5% retention. Tononi and Cirelli called this synaptic homeostasis — the brain literally downscales weak connections every night so the strong ones stand out.

---

Installation

pip install neuram

You'll need Redis. If you don't have it: brew install redis && redis-server.

Dev dependencies (pytest, fakeredis):

pip install "neuram[dev]"

---

Quick start

import asyncio
from neuram import Brain, MemoryType

async def main():
    brain = await Brain.create(redis_url="redis://localhost:6379")

    # facts go to Neocortex after consolidation
    await brain.perceive("Paris is the capital of France",
                         memory_type=MemoryType.SEMANTIC)

    # high salience → Amygdala gives it a stability boost at encoding
    await brain.perceive("I had a great meeting with Alice today",
                         memory_type=MemoryType.EPISODIC,
                         salience=0.8)

    # skills end up in Cerebellum — hard to forget once learned
    await brain.perceive("To debug Python: add breakpoint(), run, inspect locals",
                         memory_type=MemoryType.PROCEDURAL)

    # reward signal goes to BasalGanglia (dopamine-style)
    await brain.perceive("User always prefers concise answers",
                         memory_type=MemoryType.HABITUAL,
                         reward=0.9)

    # recall via hippocampal pattern completion + spreading activation
    results = await brain.recall("capital city Europe")
    for engram, score in results:
        print(f"{score:.3f}  [{engram.memory_type}]  {engram.content}")

    print(f"Working memory: {len(brain.working_memory)} / 7 slots")

    # run a manual sleep cycle — or let start_sleep_cycle() do it in the background
    stats = await brain.sleep()
    print(stats)  # {'consolidated': 2, 'pruned_wm': 0, 'pruned_ltm': 1}

    await brain.close()

asyncio.run(main())

---

A few things worth knowing

Background sleep. You probably want this running continuously:

brain = await Brain.create(redis_url="redis://localhost", sleep_interval_seconds=30.0)
await brain.start_sleep_cycle()
# your agent does its thing
await brain.stop_sleep_cycle()

Explicit forgetting. Sometimes you need to surgically remove something:

await brain.forget(engram.engram_id)

Watching memories decay. Useful for debugging or for building rehearsal schedules:

from neuram import compute_retention, time_until_forgotten

compute_retention(engram)       # 0.73 — still pretty strong
time_until_forgotten(engram)    # 4.2 hours left before it drops below threshold

Bypassing the Brain and working directly with regions:

from neuram import Hippocampus

hpc = Hippocampus()
results = hpc.pattern_complete("what did we discuss about Redis?", top_k=10)

---

Why not just use a vector DB?

You can. pgvector, Chroma, Pinecone — they're all fine for search. But they don't forget. They don't have working memory limits. They treat a user's offhand comment the same as something they've said a hundred times. They don't do anything while your agent sleeps.

neuram is for when you want memory that behaves, not just memory that stores.

---

References

• Semon, R. (1904). Die Mneme — coined "engram"
• Ebbinghaus, H. (1885). Über das Gedächtnis — forgetting curve
• Collins & Loftus (1975). Spreading activation in semantic networks
• Bliss & Lømo (1973). Long-term potentiation in hippocampus
• Tononi & Cirelli (2006). Synaptic homeostasis hypothesis
• Scoville & Milner (1957). Patient H.M. — hippocampus and declarative memory
• McCormick & Thompson (1984). Cerebellum and procedural memory

---

MIT License