Deep-tech consulting&integrated systems architecture.

PRODUCTS

Wisteria

A native BYOK AI companion built around one persistent conversation, with memory and continuity owned by the session rather than the underlying model.

The conversation never resets

Wisteria does not reconstruct a past conversation after each session reset. The conversation itself persists. One session owns its raw history, persona, long-term memory, journals, memos, reminders, and settings, allowing the same accumulated context to continue over time.

Continuity belongs to the session

The underlying AI model is a replaceable inference engine, not the owner of the conversation. Changing models or providers does not replace the session or discard its accumulated history. Conversation models and memory-compression models can also be configured independently.

Memory changes resolution, not continuity

Recent conversation remains at high resolution while older experience is compressed into progressively larger narrative units: turn summaries, short events, mid-term context, long-term patterns, and life-scale narrative. The raw conversation is preserved separately rather than being replaced by the compressed memory.

Different information uses different memory channels

Narrative memory preserves temporal flow, recurring patterns, relationships, and the context behind decisions. Facts, schedules, tasks, and execution states that should not blur under narrative compression can remain in memo and reminder channels, while raw conversation and journal records provide detailed reference points for recovery, review, and long-term recall.

Model-independent BYOK runtime

Wisteria separates the persistent session core from provider-specific interfaces. OpenAI and Anthropic can be used through independent provider adapters while the same conversation and memory remain intact. Users provide their own API access and retain control over the models used by each session.

Built for long-term use

The native Windows client includes multiple independent sessions, per-session persona and model settings, BYOK credentials, permanent raw-history storage, multi-stage memory compression, image memory, journals, memos, reminders, backup and restore, token and cost tracking, prompt-cache-aware context assembly, and a privacy-preserving event black box for error reporting.

An earlier alpha version was operated for more than a year, and the native BYOK client is used in ongoing development.

Technical Note — Long-Term Memory Architecture ↗

Anchoring — Reference Architecture ↗

Coming soon on Steam