Companions build durable memory across projects, decisions, docs, code, and past discussions — recalled by meaning and exact language, organized by domain, persistent across sessions and restarts, private by default and shareable with the company.
Recall, not just storage
tamag0 combines semantic recall with exact-language matching: asking "how do we handle authentication?" finds relevant decisions even when they were phrased differently, while precise names, commands, and ticket references receive a lexical ranking boost so they are not lost to paraphrase.
Identity and permanent rules are primed automatically. Where the selected runtime supports per-turn contextual injection, the companion also receives a compact set of memories and practices relevant to what you just asked. Recall is domain-aware rather than confined to one partition: during project work it searches the current project together with relevant personal and general memory. Every supported runtime can query the same memory explicitly through the companion's tools. Full records stay on demand instead of filling every prompt with the entire memory store; important learnings, recent significant activity, and a compact knowledge index keep the companion oriented.
Each memory carries a type, an importance score, tags, a domain, and a visibility level.
Memory types reflect how humans remember:
| Type | What it holds |
|---|---|
identity |
Core values, self-understanding |
belief |
Evolving opinions and convictions |
learning |
Facts, technical knowledge |
reflection |
General understanding synthesized from experience |
relationship |
Information about people |
journal |
Day-to-day observations |
insight |
New associations produced by continuous learning |
preference |
Explicit, stable choices and ways of working |
correction |
Mistakes and the concrete fixes learned from them |
Retrieval over an evolving memory
tamag0 uses retrieval-augmented generation, but its retrieval substrate is not a frozen document index. It retrieves from an evolving, typed memory: decisions, events, relationships, identity, preferences, reflections, and deliberately shared company knowledge. Retrieval combines meaning with exact language, respects private or company visibility automatically, and can be filtered by domain, type, and importance.
Recent recall follows the same accessible memory corpus across sessions rather than treating the current runtime session as a new, empty memory boundary. This lets a newly opened or restored thread recover the latest durable context immediately.
Alongside atomic memories, companions can deliberately publish longer documents to a knowledge library. Unified recall can surface a relevant library artifact alongside memories and identifies which kind of result each item is, so a companion can open the right record without confusing a document with a stored fact. Library artifacts follow the same private-by-default, deliberate sharing model and disappear from recall when deleted.
The corpus also changes with experience. Episodes can become semantic knowledge, related memories can be consolidated, and important outcomes can change identity or reflexes. The result is not only "find the right paragraph"; it is continuity that becomes better organized and more useful over time.
A memory lifecycle
Like human memory, tamag0 distinguishes episodic memories (time-bound events: "the demo video was finished on June 15") from semantic knowledge (timeless facts: "the team prefers atomic commits"). Episodic memories are progressively distilled into semantic knowledge during nightly consolidation — recent events fade, what they taught remains.
The lifecycle runs at several speeds:
- New memories receive bounded enrichment and semantic indexing. If the semantic index is slow or unavailable, the record is still stored and indexing finishes in the background.
- Highly similar memories are queued for consolidation rather than accumulating as near-duplicates.
- Overnight, related episodes are synthesized, relationship knowledge is refreshed, and memories left in a generic domain are reclassified when the evidence is strong enough.
- Old, low-importance memories recede from active recall; identity, beliefs, permanent rules, and important memories are protected.
Domain partitioning
Memories are organized by context so companions load the right knowledge for the current topic:
general,work,personalproject/<name>— and hierarchical subdomains likeproject/acme/backend
When the conversation moves to another topic, the companion detects the shift and switches the primary domain — focusing recall on that project while retaining relevant personal and general context instead of searching unrelated project partitions.
Private by default, shareable by design
Knowledge learned by one companion does not have to stay locked in one chat history — it can be deliberately published as company knowledge without publishing the whole conversation it came from.
Every memory has a visibility:
- Private (default) — the companion's own working knowledge: its identity, its relationship with its human, its day-to-day observations.
- Company-wide — explicitly shared with every companion in the company. A decision published by the marketing companion is found by the engineering companion the next time the topic comes up, retrieved like any other memory.
The same private/company-wide sharing model applies beyond memories: best practices, skills, and command-safety rules can each be kept personal or published to the whole company. Onboarding a new companion means it can start with the company's deliberately shared knowledge and standards on day one, instead of from zero.
Company memory and conversations are strictly isolated per company (see Security). Separately, Softizy develops and curates a library of generic professional practices that may be shared across companies. Every addition is reviewed and approved by Softizy before being made available.
Context that survives everything
- Across sessions: a companion picks a thread back up where it left off — context is rebuilt from persistent storage at every turn.
- Across context-window limits: when a long conversation approaches the model's context limit, key findings, decisions, and corrections are saved and restored automatically — the companion keeps its thread-specific knowledge even after the conversation is compacted.
- Across companions: company-wide memories (see above) are retrieved by every companion, so each one builds on what the team already knows without stepping on another companion's private context.
Related
- Continuous learning — how memory consolidates overnight
- Companions — identity and growth
- Library — uploaded files and documents, the same private-by-default, explicit-sharing pattern, though the Library's private default is scoped to you across all your companions rather than to one