— memory & learning

It remembers.
And it learns.

Pumaï isn't limited to the recent transcript. It "knows" its topics and the people it talks to over time — and stabilizes what it learns every night.

— three kinds of memory

Three memories that coexist

narrative

Who it is

Its identity, its voice, its durable facts. Injected first — the persona always leads.

conversational

What was said

From the most recent to the oldest, in three stages (below).

the person

What it knows about you

Built automatically as you talk: role, interests, recurring topics.

— conversational decantation

Three stages, from hot to cold

Conversational memory settles progressively. Nothing is lost — but nothing needlessly clutters the context.

stage 1 — live

Sliding window

Recent messages stay verbatim. When the window overflows, its oldest part is summarized and kept.

stage 2 — warm

Summarized segments

After a pause, a block of conversation is summarized and archived. The most recent ones are re-injected automatically.

stage 3 — cold

Searchable memory

Older material becomes a memory the agent retrieves on demand — never force-fed into the prompt.

Tier 1 — Sliding window · RAM raw messages; min. 12 kept; on overflow → LLM summary of the oldest half
↓ pause ≥ 60 min — the session closes and archives
Tier 2 — Segments · PostgreSQL LLM-summarized block; the 3 most recent segments are auto-re-injected
↓ beyond the 3 recent — vectorized
Tier 3 — Vector RAG · pgvector HNSW cosine 1024-d index, partitioned per profile; never auto-injected — retrieved on demand via search_memory
memory stratification — from hot (RAM) to cold (vectors)
— the nightly dream

Every night, it consolidates

Pumaï reviews its recent activity, detects recurring patterns, promotes the durable into long-term memory, and forgets the noise. It even prepares the next session — what to pick up, what to ask, what to avoid.

Inputs — ~03:00, one job per profile last 7 days of notes · life journal · recent segments · missions
↓
Consolidation LLM recurring patterns · hypotheses · continuity · prep for the next session
↓
Promote → durable memorypattern ≥ 3× / 7 days, FIFO 30
Writes the "compass"uncertainties, tone, topics to deepen
Enriches the person's profilerole · interests · patterns
↓ then hygiene
Forgets the noisepurge notes > 30 days · vectorize segments (if RAG enabled)
the nightly dream — consolidate without retraining the model · fully semantic sorting
— stable identity

The persona always leads. Past memories are silent background: the current request stays priority, and the assistant never rehashes its old obsessions.

— access

A use case in mind?

Tell us your use case — individual reply within 48h.

contact us →