SpikeMem by PsycorAI

Long-term memory for AI agents, built from spiking neurons and synapses.

SpikeMem gives your agent a memory that runs on your own machine and needs no LLM to write or recall memories.

A small spiking encoder turns each memory into a sparse code, and recall runs through synapses, not a vector database. Saving a memory costs no tokens. Recall is built to run on an ordinary laptop CPU, with no GPU.

Each store is one file you own. Copy it, back it up, move it to another machine. You choose which agents can read a store and which can write to it. Any agent that speaks MCP can use it.

Forgetting is real. When you delete a memory it leaves the file and the index, and you get a receipt. The store then matches one that never held it.

We're not the first here: Hebbian, associative and spiking ideas have all been applied to agent memory before, including HippoRAG, HeLa-Mem, SynapticRAG and EMBER. But SpikeMem is the only memory system for AI agents we know of that needs no LLM, has no transformer embedder, and has a spiking, synaptic recall path. We did use transformers and LLMs to train it, but none of them run when you use it.

And we haven't published benchmark numbers yet. When we do, they'll come from a held-out test set, with every system read by the same model and the caveats listed beside them.