Agent memory with zero LLM in the loop
shodh-memory forms, stores, and retrieves memories without a large language model anywhere in the loop. Extraction, ranking, and recall are deterministic code; small frozen models handle perception only. The result is memory you can audit, that never invents what it stores, and that never leaves the machine.
What "no LLM in the loop" actually means
It is a precise claim, not a slogan. The difference is where the language model sits — and whether it is allowed to decide what you remember.
# Most memory systems put an LLM inside the loop
ingest -> [ LLM extracts "facts" ] -> store -> [ LLM decides recall ] -> agent
^ can fabricate ^ opaque, unauditable
whatever the model invents or distorts is baked into storage as ground truth.
# shodh-memory keeps the loop deterministic
ingest -> [ rules + frozen NER / embeddings ] -> store (with provenance)
^ perception only, no generation ^ every memory -> a source
-> [ deterministic ranking + decay ] -> agent
^ inspectable, reproducible
the language model, if any, lives in the agent on top -- never in the memory.Why it matters
Nothing is fabricated into storage
A system that uses an LLM to extract memories can write down things that were never said — a hallucination, stored as if it were ground truth. shodh-memory does not generate; it records. Every memory traces back to the source it came from, so you can always check where a belief originated.
Deterministic and auditable
What gets stored, strengthened, or surfaced is decided by inspectable rules with documented constants — not a model's opaque judgement. The same input produces the same memory every time, so you can audit exactly why anything surfaced, and reproduce it.
Nothing leaves the device
No cloud, no API keys, no external calls. Memory is formed and queried entirely on the machine. Nothing is transmitted to, cached by, or retained by a third party — there is no remote copy to lose.
No extraction attack surface
When an LLM reads untrusted text to build memory, that text can steer what gets stored — a poisoning vector hidden in ordinary input. Deterministic extraction has no prompt to hijack, so a whole class of memory-poisoning attacks simply has nowhere to land.
Where these properties tend to matter
We build the memory layer. Whether these properties are decisive for your system is your call — but they tend to matter most in places like these.
Regulated & audited environments
When you have to answer “why does the system believe this?”, memory with provenance and reproducible behaviour is auditable by construction, rather than after the fact.
Air-gapped & edge deployments
Where there is no network — or none is permitted — memory that forms and recalls entirely on-device keeps working through disconnection and restarts.
Security-sensitive agents
When the text an agent ingests cannot be trusted, an extraction path with no LLM to hijack removes the prompt-injection route into long-term memory.
Personal & on-device assistants
When memory holds personal or health data, keeping it on the machine means it is never somewhere else to be exposed.
Go deeper
Memory the agent can use — without handing the agent control over what it remembers.
One binary, one command, no LLM in the loop.